Author: Gabriel M.

The competitive dynamics in AI between the US and Chinese ecosystems have evolved in ways that the export control framework of 2022 and the broader US AI strategy did not fully anticipate. DeepSeek’s R1 release in early 2025 demonstrated that Chinese AI labs could produce frontier-quality models with substantially less compute capital than US labs had been deploying. Alibaba’s Qwen model family has continued aggressive development with strong open-source distribution, producing capable models that operate at scale across Chinese cloud infrastructure and that have been adopted globally as open-weight alternatives to closed proprietary models. ByteDance’s various AI deployments — both within the consumer applications (TikTok, Douyin) and through the broader Doubao AI initiatives — have demonstrated vertically integrated AI deployment at the scale that only the largest consumer technology companies can match.

The result is an AI competitive environment in 2026 where the Chinese AI ecosystem has demonstrated capabilities that compete credibly with Western alternatives, where the open-source distribution strategy that Chinese labs have prioritised has created adoption dynamics that affect the global AI economics, and where the export control regime that was designed to constrain Chinese AI development has produced effects that are more complex than the simple containment narrative implied. Understanding what has actually happened and what it means for the broader AI investment environment requires looking at the specific Chinese AI developments and the structural dynamics that have produced them.

The DeepSeek Inflection and What It Actually Showed

The DeepSeek R1 release in early 2025 was the most consequential public moment in Chinese AI development to date. The model demonstrated frontier-quality reasoning capabilities that competed credibly with OpenAI’s o1 release, achieved through training approaches that the DeepSeek team disclosed in technical papers that the broader AI research community could evaluate. The reported training cost — substantially below the costs that US AI labs had been incurring for comparable model capabilities — produced significant market and policy reactions.

The honest technical assessment of what DeepSeek demonstrated is more nuanced than the initial market reaction implied. The training efficiency improvements that DeepSeek reported reflected legitimate algorithmic and engineering innovations that the broader research community has been able to validate and apply. The reported training costs, however, captured only specific portions of the actual development costs (not including the broader research investment, the prior model development that supported R1’s specific advances, or the infrastructure that supported the training). The full economic picture of Chinese AI development is more expensive than the headline R1 training cost figure suggested.

The broader strategic implication, however, was substantial regardless of the specific cost accounting. The demonstration that frontier-quality models could be produced by Chinese labs operating outside the export control regime — using domestic Chinese semiconductors (Huawei Ascend, the various other Chinese AI chip alternatives) and various indirect access to Western capabilities — challenged the foundational assumption of the US export control strategy. The framework that assumed export controls could meaningfully constrain Chinese AI development was undermined by the empirical evidence of Chinese labs producing competitive capabilities despite the restrictions.

The Alibaba Qwen Open-Source Strategy

Alibaba’s Qwen model family has executed the most consequential open-source AI strategy of the past two years. The Qwen models have been released with permissive open-source licensing, with multiple model sizes covering the breadth of deployment scenarios, and with continued aggressive development pace that has produced model generations at frequent intervals. The Qwen models have been adopted globally as open-source alternatives to Meta’s Llama family, with substantial usage in research, in production deployments at companies that prefer open-source models, and in the broader AI development ecosystem.

The strategic logic for Alibaba is multi-layered. The open-source distribution accelerates Qwen adoption beyond what closed proprietary distribution could achieve, which produces ecosystem development that supports Alibaba’s broader AI infrastructure business through Aliyun (Alibaba Cloud). The international Qwen adoption provides Alibaba with brand recognition and developer mindshare in markets where Chinese AI alternatives had not previously been considered. The competitive positioning against US proprietary alternatives benefits from the open-source distribution providing a credible alternative to enterprises evaluating their AI vendor commitments.

The broader AI infrastructure competitive dynamics are affected by Qwen’s open-source distribution in ways that have implications for the Western AI providers. Alibaba Cloud’s positioning as the primary Qwen deployment infrastructure provides competitive differentiation in the Asia-Pacific region where Alibaba’s broader cloud business operates. The Qwen models running on competing cloud platforms (AWS, Azure, Google Cloud through their multi-model integration) provide Alibaba with influence beyond its direct cloud customer base.

The ByteDance Vertical Integration

ByteDance’s AI deployment strategy operates through a different model than Alibaba’s open-source distribution or DeepSeek’s pure-research positioning. ByteDance has integrated AI capabilities deeply into its consumer applications (TikTok and Douyin globally, the various Chinese-specific applications), with AI used for content recommendation, video generation, content moderation, and the various other capabilities that the consumer products require.

The Doubao AI initiatives have built ByteDance-specific foundation model capabilities that compete with the international AI providers in the Chinese market and that have been deployed for various consumer and enterprise applications. The vertical integration that ByteDance has achieved — controlling the consumer applications that deploy AI, the underlying AI capabilities, and the broader infrastructure that supports both — represents a different competitive position than either pure-AI labs like OpenAI or pure-cloud providers like Alibaba.

The TikTok regulatory situation has been one of the most important political dynamics affecting ByteDance’s broader AI positioning. The various TikTok divestiture and regulatory pressures across multiple countries have created uncertainty about ByteDance’s ability to maintain its consumer application footprint, which affects the broader vertical integration thesis. The outcome of the various TikTok regulatory situations will affect ByteDance’s positioning across multiple AI deployment dimensions.

The Export Control Regime and Its Actual Effects

The US export control regime targeting AI semiconductors has been one of the most consequential industrial policy initiatives of the post-2022 period. The controls have substantially restricted Chinese access to leading-edge Nvidia GPUs and to the equipment needed to manufacture advanced semiconductors domestically. The intended effect was to constrain Chinese AI development by limiting access to the compute that frontier AI training requires.

The actual effects have been more complex than the simple containment framework anticipated. The Chinese AI labs have continued to produce competitive capabilities despite the export controls, partly through domestic chip alternatives (Huawei’s Ascend chips have improved substantially), partly through algorithmic and training efficiency improvements that have reduced the compute requirements for specific model capabilities, and partly through various indirect access mechanisms that have not been fully closed by the export control framework.

The semiconductor supply chain concentration has interacted with the export control regime in ways that have produced specific effects. The export controls have constrained Chinese access to the most leading-edge capabilities but have not prevented the broader Chinese AI capability development at scales that compete with Western alternatives.

The honest assessment of the export control regime is that it has produced some delay in specific Chinese AI capability development and has created friction that affects Chinese AI economics, but it has not produced the structural containment that the policy framework anticipated. The Chinese AI ecosystem has adapted to the constraints in ways that the policy framework’s designers did not fully model. The implications for the broader US-China AI competitive dynamics are that the export control regime has affected the trajectory at the margin without fundamentally changing the structural competitive picture.

The Open-Source AI Economics Implication

The Chinese AI ecosystem’s emphasis on open-source distribution has implications for the global AI economics that affect Western AI providers. The availability of capable open-source models (Qwen, the various other Chinese open-source releases, supplemented by Meta’s Llama family) creates competitive pressure on the closed proprietary model pricing that OpenAI, Anthropic, and the other Western AI providers have established.

The specific competitive dynamics include the enterprise customer segment that increasingly evaluates open-source alternatives for use cases where the closed proprietary capabilities do not provide proportional value, the AI infrastructure provider segment that benefits from being able to offer open-source models alongside closed alternatives, and the developer ecosystem that has integrated open-source models for various applications where the open-source flexibility provides specific advantages.

OpenAI’s monetisation challenges include the open-source competitive pressure as one of the structural factors affecting its business model. The closed proprietary AI providers’ ability to maintain pricing power depends partly on the open-source alternatives not closing the capability gap that justifies the proprietary pricing premium. The Chinese open-source releases have been one of the most significant contributors to closing that gap, alongside Meta’s Llama development that has been the primary Western open-source contribution.

The Investment Implications

For investors evaluating exposure to the AI investment cycle: the Chinese AI competitive picture affects the overall investment thesis in ways that the simple US-vs-China framing does not capture. The closed proprietary Western AI providers face structural competitive pressure from both Chinese open-source alternatives and from the broader open-source ecosystem that includes Western contributions (Meta’s Llama family) alongside the Chinese contributions.

The semiconductor companies that benefit from AI infrastructure demand have complex exposure to the Chinese AI dynamics. Nvidia’s revenue has been affected by the export control regime but has been substantially supported by Western demand that has dwarfed the constrained Chinese segment. The Chinese chip alternatives (Huawei Ascend, the various other Chinese AI semiconductors) compete primarily in the Chinese market without yet substantially affecting the global semiconductor competitive picture.

The cloud providers’ AI strategies are affected by the Chinese AI dynamics in different ways. Google’s broader AI positioning includes the question of how to compete with Chinese AI providers in the broader international markets where both compete. AWS and Azure’s positioning depends partly on the relative attractiveness of the various AI models they integrate, which includes the open-source Chinese alternatives alongside the Western closed proprietary models.

For the specific Chinese AI companies that are publicly investable (Alibaba primarily, with Tencent and several others having different specific exposures), the AI positioning provides upside that the broader Chinese equity environment continues to underprice. The Chinese macro picture’s broader challenges have affected Chinese equity valuations, which means the AI positioning that companies like Alibaba have achieved is not fully reflected in current valuations.

The Honest Strategic Assessment

The Chinese AI capability development has been more rapid and more impressive than the policy framework that was designed to constrain it anticipated. The structural competitive picture has evolved into a multi-polar AI environment where the US, Chinese, and broader international AI ecosystems all have meaningful capabilities and where the competition operates across multiple dimensions (consumer applications, enterprise services, open-source distribution, semiconductor infrastructure, regulatory frameworks) that produce different competitive winners in different segments.

The implications for global investors are that AI exposure should be evaluated as a more complex multi-dimensional investment theme than the simple “Western AI vs Chinese AI” framing implies. The specific company positions, the open-source vs proprietary competitive dynamics, the semiconductor infrastructure exposures, and the regulatory and political risks all affect the appropriate positioning for AI investment exposure in ways that require sophisticated analysis rather than broad sector allocation.

The honest position is that Chinese AI has become a serious competitive force that affects the global AI investment environment in real ways, that the US export control regime has been less effective at containment than the policy framework anticipated, and that the open-source distribution strategy has produced competitive dynamics that benefit the broader AI ecosystem at the expense of closed proprietary AI economics. The next several years will continue to test the various competitive positioning, but the structural picture has evolved into one where the global AI competition is genuinely multi-polar rather than US-dominated.

The Mental Model for Understanding What a Real Capability Shift Looks Like

Shane Parrish’s work on mental models returns consistently to one meta-principle: the frame you use determines what you can see. The US technology policy establishment used an export control frame to evaluate Chinese AI development — measuring capability gaps by access to frontier chips and assuming that closing the chip gap would take years. That frame produced a systematic forecast error because it was measuring the wrong variable. Capability in AI systems is not primarily a function of chip access at the frontier. It is a function of algorithmic efficiency, training data quality, engineering organizational capacity, and the speed of iteration cycles. China’s leading AI labs have demonstrated advantages in at least three of those four dimensions. The chip gap is real. The capability gap it was supposed to produce is not.

DeepSeek’s training efficiency result — frontier-competitive performance at a small fraction of the compute cost of comparable Western models — is the key data point that broke the export control frame. If capability is a function of compute, controlling compute controls capability. If capability is a function of algorithmic efficiency, controlling compute slows capability development without stopping it. DeepSeek demonstrated empirically that the second model is closer to correct. The policy frame that justified the export controls was not wrong about the importance of compute. It was wrong about whether compute is the binding constraint on capability at the current stage of AI development.

Parrish’s second-order thinking principle asks: if DeepSeek’s efficiency is real and replicable, what happens next? The immediate implication is that enterprise AI adoption becomes a competition between Western and Chinese models on price and performance, not just between Western models. An enterprise evaluating AI infrastructure in 2026 has access to Qwen, DeepSeek, and ByteDance’s models as genuine alternatives to GPT-4 and Claude, with inference costs that are dramatically lower at comparable capability levels. That competition has not yet fully reached Western enterprise procurement, but it is in the pipeline. The enterprises that are tracking this development are getting price leverage in their AI infrastructure negotiations that enterprises ignoring it are not.

The open-source distribution strategy is the second dimension of the catch-up that the export control frame missed entirely. Alibaba’s Qwen open-source release strategy converts Chinese AI capability into global distribution at near-zero marginal cost. Every developer who builds an application on Qwen is a distribution node for Chinese AI capability that exists entirely outside the export control regime. The capability spreads through open-source adoption rather than through hardware supply chains. Export controls can slow hardware access. They cannot slow open-source model adoption.

The mental model correction that this article’s evidence requires is replacing “China is catching up” with “China has arrived at a different architectural approach that produces competitive results more efficiently.” Those are not the same statement. “Catching up” implies the Western frontier is still clearly ahead and China is reducing a gap. “Different approach producing competitive results” implies the frontier is now contested across multiple architectural strategies, and the winner of the next phase is genuinely unclear. The capital rotation away from earlier AI certainties reflects this uncertainty: institutional capital that was confident in the Western AI monopoly thesis is repricing its confidence downward as the competitive evidence accumulates. Prediction markets on Chinese AI market share in enterprise deployments by 2027 are now pricing a meaningful Chinese position — not because of optimism about China, but because the empirical evidence of the capability has become too clear to price away.

The Hyperliquid Liquidity Provider (HLP) vault has emerged as one of the most strategically interesting product innovations in crypto over the past two years. The vault allows any participant to deposit USDC and become a fractional participant in Hyperliquid’s market making operations on the protocol’s perpetual futures exchange, sharing in the profits and losses that market making activity generates. The HLP vault has attracted deposits in the billions of dollars, has generated consistent positive returns for depositors across most reporting periods, and has demonstrated something genuinely new about how on-chain market making can be democratised and scaled.

The broader Hyperliquid story has been one of the most discussed in crypto over the past year — the protocol has grown to capture substantial share of perpetual futures trading volume, the HYPE token has performed strongly, and the various strategic considerations including the potential public market listing have generated significant attention. The HLP vault specifically deserves attention because it represents a different kind of innovation than the trading platform itself: a mechanism that converts market making — historically the domain of sophisticated proprietary trading firms — into a structured product that public depositors can participate in.

Understanding what the HLP vault actually does, how the economics work in practice, and where the structural risks sit provides important context for evaluating both the specific Hyperliquid investment thesis and the broader question of whether the HLP model can be replicated or whether it represents a genuinely unique innovation.

What HLP Actually Does

The HLP vault operates Hyperliquid’s market making strategies on the protocol’s perpetual futures exchange. Market making involves continuously posting bid and ask quotes across the trading pairs, capturing the spread between buy and sell prices when trades execute, managing the resulting inventory exposure through hedging across related instruments, and absorbing the temporary directional risk that comes from inventory imbalances before the imbalances can be cleared.

The strategies that HLP executes are sophisticated quantitative trading approaches that have been developed and refined by the Hyperliquid team. The specific details of the strategies have not been fully disclosed (which is appropriate for proprietary trading approaches that could be replicated or front-run if fully transparent), but the broad outline involves posting two-sided quotes across the perpetual futures markets that Hyperliquid supports, dynamically adjusting the quoted prices based on inventory positions and broader market conditions, and hedging the resulting risk through positions in related instruments.

The economic value proposition for HLP depositors is direct participation in the market making revenue that Hyperliquid generates without requiring the operational sophistication, capital scale, or technical infrastructure that running independent market making operations would require. A depositor effectively buys fractional exposure to Hyperliquid’s market making book at the cost of accepting the strategy risk that the team’s approach involves.

The Returns Profile and What It Reveals

The HLP vault has produced annualised returns that have generally been in the 15-35 percent range across reporting periods, varying with market conditions, trading volume on the protocol, and the specific positions that the market making strategy has held during different periods. The returns have been positive in most periods but have included some negative periods during specific market dislocations where the strategy positioning produced losses that the broader profitable activity could not fully offset.

The honest reading of the returns data is that they reflect genuine market making profitability that has been consistent enough to attract substantial deposits while being variable enough to require depositor understanding of the underlying risk profile. The returns are not predictable in the way that yield-bearing stablecoin returns are predictable; they are market-condition-dependent in ways that any market making strategy is dependent on the trading activity and market dynamics that produce the spread capture.

The comparison to alternative on-chain yield strategies is instructive. The stablecoin yield alternatives generally produce more predictable but lower returns. The DeFi lending alternatives provide returns that depend on borrowing demand. The HLP vault returns are higher than most alternatives on average but with substantially more variance and with structural risks that are specific to market making activity rather than to credit or rate exposure.

The depositor base for HLP has grown substantially as the returns track record has accumulated. The depositors include both crypto-native individuals seeking yield on their USDC balances and institutional participants who have evaluated HLP as an alternative to other crypto yield opportunities. The institutional participation has been particularly meaningful because it represents external validation of the strategy and the operational infrastructure that supports it.

The Structural Risks Worth Understanding

The HLP vault’s risk profile is genuinely different from the broader DeFi yield landscape because the underlying activity (market making) involves specific risks that are different from the credit, rate, and protocol risks that affect other DeFi yield products. The structural risks that depositors should understand include strategy risk (the specific approaches that the Hyperliquid team employs may produce losses in market conditions that differ from those that the strategies are calibrated for), execution risk (the operational infrastructure that runs the strategies needs to perform reliably across market conditions), and liquidity provision risk (during periods of severe market stress, market makers can face large directional moves that produce concentrated losses).

The specific market making strategy risks include the possibility of inventory positions that cannot be hedged effectively during fast-moving market conditions, the impact of large counterparty positions that may create non-typical order flow patterns, and the dynamic adjustment of strategies as market conditions change in ways that may not be optimal for all market environments. These risks are inherent to market making rather than specific to HLP, but they are risks that public market making structures expose depositors to in ways that traditional yield products do not.

The Hyperliquid team has developed risk management approaches that include position limits, hedging requirements, and the broader strategy oversight that maintains the operational integrity of the market making activity. The transparency of the strategy operations (visible on-chain through the position data and trading activity) provides depositors with visibility into the activity that they are participating in. The combination of risk management and transparency has supported the trust that has allowed deposits to grow to substantial scale.

The Protocol Revenue Architecture and HLP’s Role

The HLP vault is one component of the broader Hyperliquid protocol revenue architecture that supports the HYPE token economics. The protocol generates revenue from trading fees, from the funding rate mechanism that perpetual futures use, and from the various other operational components of running a perpetual futures exchange. The HLP vault participates in the market making revenue specifically, while other revenue components flow to the broader protocol treasury and to HYPE token holders through the various distribution mechanisms.

The strategic positioning of HLP within the broader protocol revenue is important because it aligns the depositor interests with the protocol’s broader success. HLP depositors benefit from substantial trading activity on the protocol (which produces market making opportunities), from the protocol’s ability to attract liquidity from other sources (which makes the market making strategies more effective), and from the broader ecosystem development that supports the protocol’s competitive positioning.

The broader DEX value capture dynamics apply in interesting ways to HLP. Where other DEX protocols have struggled with the question of how token holders capture the trading volume value, Hyperliquid’s architecture provides multiple mechanisms for value capture (HYPE token economics, HLP vault participation, the broader protocol revenue). The specific mechanism that depositors use for value capture (HLP vault for market making revenue, HYPE token holding for broader protocol revenue) depends on their preferences and risk tolerance.

The Replicability Question

A natural question about the HLP vault innovation is whether other perpetual futures DEXes can replicate the model and whether the HLP advantage represents a sustainable competitive moat for Hyperliquid. The honest assessment is that the model is replicable in concept but is difficult to execute at the same level of sophistication that Hyperliquid has achieved.

The barriers to replication include the specific quantitative trading capability that supports the market making strategies (which depends on the team’s expertise rather than just the protocol infrastructure), the depositor trust that allows substantial capital to be committed to the vault (which builds over time based on demonstrated performance), and the broader ecosystem development on the protocol that makes the market making opportunities attractive (which depends on the protocol’s broader success).

Several other DEXes have launched vault-like products that attempt to provide similar exposure to market making revenue, but none have achieved the scale or the operational sophistication that HLP has demonstrated. The specific advantages that Hyperliquid has — the first-mover positioning, the substantial trading volume that supports market making opportunities, the operational track record that has built depositor confidence — represent a competitive moat that other protocols would need to overcome to provide equivalent products.

The Investor Considerations

For investors evaluating HLP vault exposure: the returns are attractive but require understanding of the underlying market making risk profile, the historical performance is encouraging but is not a guarantee of future performance, and the deposit decision should be sized appropriately for the risk tolerance of the investor’s broader portfolio.

The structural advantages of HLP — substantial scale, professional strategy execution, transparent operations, integration with one of the leading perpetual futures DEX protocols — make it one of the more credible on-chain yield opportunities for investors willing to accept the market making risk profile. The structural risks — strategy-specific exposure, the possibility of stress period losses, the dependence on continued protocol success — should be priced into the investment decision rather than ignored.

For investors evaluating the broader Hyperliquid investment thesis (HYPE token exposure, perpetual futures DEX category exposure, the various other components of the protocol ecosystem): the HLP vault is one component of a multi-faceted investment thesis that depends on the continued success of the broader protocol. The integration of HLP into the broader protocol revenue architecture means that HLP success and HYPE token success are correlated, which has implications for portfolio construction across the various Hyperliquid-related exposures.

The honest position is that the HLP vault represents one of the more innovative product structures in crypto, that the returns have validated the model at scale across multiple market conditions, and that the structural risks are real but manageable for investors who understand the underlying market making dynamics. The category of on-chain market making vaults that HLP has effectively pioneered will likely produce additional entrants and variations over the next several years, but the specific HLP advantages position it as the category leader for the foreseeable future. The broader implication is that on-chain market making at scale is feasible, that public participation in market making revenue can be structured effectively, and that the crypto category has produced product innovations that have no direct equivalent in traditional finance — which is exactly the kind of structural innovation that crypto’s institutional adoption thesis has long anticipated.

The Mechanism Observed Closely: What HLP Is Actually Doing at the Trade Level

John McPhee’s method is to get close enough to the subject that the mechanism becomes visible. Abstract descriptions of what something does are less useful than precise accounts of what actually happens. Applied to HLP, the mechanism worth understanding is not the yield number — it is the specific trade that produces the yield, the specific risk that produces the loss tail, and the specific structural advantage that allows HLP to perform a function that traditional market-making operations cannot perform in the same way.

HLP provides liquidity to Hyperliquid’s order book. When a trader opens a perpetual position on Hyperliquid, HLP is on the other side of that trade if no other counterparty is available at the required price. HLP earns the spread between bid and ask, collects a portion of funding rates when positions are directionally skewed, and absorbs the mark-to-market loss when the positions it holds move against it. The vault’s positive expected value depends on the spread income and funding rate collection exceeding the mark-to-market losses over time. In liquid, mean-reverting markets, this works reliably. In trending markets with large directional positions, it does not.

MEV extraction dynamics on Ethereum provide a useful contrast for understanding what Hyperliquid’s architecture is avoiding. Traditional DeFi market makers operating on AMM-based protocols face a specific category of adversarial extraction: sandwich attacks, front-running, and JIT liquidity provision that captures the spread without bearing the inventory risk. Hyperliquid’s order book model, combined with its validator set and block structure, is designed to minimise this extraction. HLP benefits from this design because the spread it earns is not being systematically captured by faster participants operating on the same infrastructure.

The collapse of FTX created the specific market condition that made Hyperliquid’s growth trajectory possible. The institutional market-making infrastructure that FTX had built — its proprietary trading arm, its liquidity provisioning relationships, its cross-exchange arbitrage operations — was removed from the DeFi perps market simultaneously. HLP entered into a market where the dominant competitor had disappeared, where retail demand for perpetual exposure had not, and where no alternative centralised venue had yet established the same level of trust that FTX had prior to its collapse. The timing was structural, not coincidental.

The Maker protocol risk model offers a relevant comparison for thinking about HLP’s structural risk. Maker’s stability fee and collateralisation ratio system is designed to ensure that the protocol remains solvent under adverse price conditions. HLP has a different but analogous risk management challenge: it must ensure that the vault’s collateral remains sufficient to cover its open positions under adverse conditions, without the liquidation mechanism that Maker uses. The socialised loss mechanism — where losses are spread across all vault depositors — is HLP’s equivalent of Maker’s stability mechanism. It works until the position that generates the loss is large enough to exceed the vault’s buffer.

stablecoin B2B payment infrastructure is relevant to HLP’s collateral base. The vault’s deposits are denominated in USDC. As stablecoin B2B infrastructure matures and more institutional capital flows through stablecoin-denominated channels, the depth of USDC liquidity that HLP can access improves. More depositors means more capacity to absorb the large directional position that is the vault’s primary risk scenario. The relationship between stablecoin infrastructure maturity and HLP’s risk capacity is indirect but real.

Privacy infrastructure for on-chain trading represents a category of technical development that could affect HLP’s competitive position. If ZK-enabled private order books become viable at scale, the information advantage that Hyperliquid’s transparent order book provides to the market-making function changes. Private order flow is both an opportunity — HLP could benefit from being the counterparty to informed private flow — and a risk, because the signal extraction that helps HLP manage inventory becomes harder when orders are not visible. The timeline for this to become relevant is measured in years, not quarters.

HLP is a genuine financial innovation. Understanding it precisely — rather than through the lens of the yield number alone — is the prerequisite for evaluating whether the risk-reward is what the depositor base believes it to be.

Mental Models for Evaluating DeFi Vault Economics: What HLP Actually Tells You

Shane Parrish’s mental model library includes a specific framework for evaluating complex financial mechanisms: follow the incentives, not the description. The description of a financial product tells you what the designers want you to think about it; the incentive structure tells you what the participants are actually optimizing for. Applied to Hyperliquid’s HLP vault, the description is “public market making infrastructure that earns fees by providing liquidity to the order book.” The incentive structure reveals something more specific: the HLP vault is a mechanism that allows passive capital to participate in the market-making activity that was previously available only to sophisticated operators with direct exchange access, in exchange for bearing the directional risk that market making in perpetual futures entails when the order flow is imbalanced.

Parrish’s second-order thinking framework asks: what happens next after the first-order effect? The first-order effect of the HLP vault is that it provides liquidity to the Hyperliquid order book, enabling tighter spreads and better execution for traders. The second-order effect is that by making market-making accessible to passive capital, Hyperliquid has created a participant class that has a financial interest in the platform’s continued volume growth — the HLP depositor is not just a liquidity provider, but a stakeholder whose vault return is directly correlated with the exchange’s success. This is structurally different from the relationship between a liquidity provider and a traditional exchange: the traditional LP provides liquidity in exchange for a fee but has no ownership claim on the exchange’s success, while the HLP depositor’s return is denominated in the exchange’s own economic activity. The second-order effect is that this creates a natural advocacy dynamic among HLP depositors that an exchange cannot easily manufacture through marketing.

The inversion mental model — ask what would have to be true for this to fail rather than for it to succeed — identifies the HLP vault’s primary risk as the directional exposure risk during correlated market stress. When perpetual futures markets experience large directional moves — the kind where most active traders are positioned the same way — the market maker’s counterparty risk concentrates: the vault is on the other side of the position that everyone is taking in the same direction. The vault’s risk management through position limits and funding rate adjustments is the mechanism that is supposed to contain this risk, but no mechanism can eliminate the fundamental exposure of a market maker to correlated order flow. Enterprise AI risk management frameworks face the same second-order thinking problem: the first-order benefit (AI improves decision speed) creates a second-order risk (AI-assisted decisions correlate across users, concentrating systemic risk in ways that pre-AI decision frameworks did not produce). The HLP vault’s correlated stress exposure and the AI decision correlation risk are both second-order effects that the product description does not emphasise but that the incentive structure reveals as load-bearing.

Parrish’s map-and-territory framework — the distinction between the model of the thing and the thing itself — is the most relevant lens for evaluating the HLP vault’s published performance figures. The performance figures are the map; the underlying market-making activity is the territory. The map is accurate for the period it covers, but the territory changes: market conditions that were favorable to market-making in 2025 (sufficient spread income relative to directional risk, manageable funding rate volatility) may be less favorable in 2026 as the market structure evolves and more sophisticated capital competes for the same spread income. Performance reporting without behavioral context is the most common map-territory confusion in DeFi: the published APY is the map, and the territory is the specific market conditions under which that APY was generated and the probability that those conditions persist. On-chain private credit yield faces the same map-territory problem at the lending layer: the stated yield is the map, and the territory is the credit quality distribution of the borrower pool under the market conditions that will actually determine repayment. Berachain’s BGT emission model is the system-level map that sets the incentive context within which Hyperliquid-adjacent protocols on the Berachain ecosystem operate — the BGT directed to productive liquidity pools is the territory signal that reveals where the incentive structure is actually concentrating capital rather than where the marketing materials claim it is concentrating. Prediction markets on Hyperliquid’s TVL and fee revenue through end-2026 are pricing continued growth from the current base — which Parrish’s second-order framework reads as the market pricing the first-order effect without adequately pricing the correlated stress exposure that the second-order analysis reveals.

Nate Silver’s probabilistic framing separates two distinct categories: what a data event actually measures, and what market commentary treats it as measuring. Q2 2026 earnings season will be described — in advance, in real time, and in retrospect — as a referendum on AI capex ROI, consumer resilience, margin sustainability, and the macro growth trajectory. Most of those referendum descriptions are misleading. A single earnings season is a noisy data point that selectively confirms whichever prior the analyst held before the season started, because the range of earnings outcomes consistent with any given macro regime is wide enough that almost any result can be framed as evidence for the prevailing narrative. The useful signal is in the specifics: capex guidance revisions, margin commentary on AI-related expenditure, and forward billings rather than reported revenue. Microsoft’s stock underperformance against Alphabet and Amazon is a useful calibration benchmark: the market is already differentiating between AI capex allocators on outcomes, not just on inputs. The earnings season’s most informative outputs will be the companies where that differentiation becomes sharper — where the AI investment thesis is confirmed or challenged at the specific-revenue level rather than at the narrative level that dominates pre-season commentary.

The Q2 2026 earnings season that begins in mid-July is the most consequential reporting cycle of the year for evaluating where the US equity market actually stands relative to the assumptions embedded in current valuations. The narrative drivers that have supported the major equity rally — AI capex justified by AI revenue growth, ex-technology corporate earnings resilience supporting market breadth, and corporate guidance language signaling sustained confidence — will all be tested through specific revenue, margin, and forward-looking commentary that the reporting period will produce.

The better analytical framework for the season is three specific questions whose answers actually determine portfolio positioning — rather than the broader noise of headline beat-and-miss statistics that dominate most earnings coverage. Each question reveals something about whether the structural assumptions that support current valuations are holding or weakening, and each has implications that extend well beyond the specific quarter being reported.

Question One: Is Mega-Cap AI Revenue Finally Pacing With Capex?

The most consequential question for the broader US equity market is whether the AI capital expenditure cycle that has dominated the hyperscaler narrative is producing the AI revenue growth that justifies the capital deployment. The mega-cap technology companies — Microsoft, Google, Meta, Amazon, and now several others — have collectively committed several hundred billion dollars in AI infrastructure capex over 2025 and 2026, and the equity valuations of these companies implicitly assume that the capex will produce AI revenue growth at scales that justify the investment.

The AI data center power buildout represents the most visible expression of this capex commitment, but the strategic question is whether the AI services running on that infrastructure are generating revenue at the rates that the capex pace implies. The historical pattern in technology capex cycles is that the infrastructure investment leads revenue by 12 to 24 months — the capex is deployed first, the revenue follows as services scale. The question for Q2 2026 is whether the revenue growth that should follow the 2024-2025 capex commitments is arriving at the pace required to validate the investment thesis.

The specific data points to watch include Microsoft Azure AI services revenue growth rate, Google Cloud AI-related revenue commentary, Amazon Bedrock revenue disclosure (which the company has been somewhat reluctant to disaggregate from broader AWS metrics), and Meta’s commentary on AI-driven advertising revenue improvements. The aggregate signal across these disclosures will reveal whether the AI revenue story is meeting, exceeding, or disappointing relative to the implicit expectations built into current valuations.

AWS specifically faces important questions about whether its competitive positioning in AI infrastructure is improving relative to Azure and GCP, and the Q2 reporting will provide updated evidence about the relative growth rates and customer momentum across the three hyperscalers. The dispersion across the cloud providers matters as much as the aggregate AI revenue story because the equity implications differ significantly depending on which providers capture market share.

Question Two: Is Ex-Technology Corporate America Producing Organic Growth?

The breadth question for the US equity market has been ongoing through 2025 and 2026: while mega-cap technology has driven the headline index returns, the broader equity market has shown more modest performance and more uncertain fundamentals. The Q2 earnings season will provide updated evidence about whether corporate earnings outside the mega-cap technology sector are growing organically — supported by revenue growth and operating leverage — or whether the growth is increasingly buyback-driven and dependent on the financial engineering that record buyback activity has supported.

The specific sectors to watch include financials (where banks’ net interest margin commentary will reveal whether deposit competition is pressuring earnings as the Fed cutting cycle proceeds), industrials (where manufacturing earnings should reflect any signs of capex acceleration outside the AI infrastructure story), consumer discretionary (where the resilience or weakening of consumer spending will be revealed in retail, restaurants, and travel earnings), and healthcare (where the GLP-1 weight loss drug economics, drug pricing pressures, and managed care utilisation trends will be visible).

The aggregate question is whether the S&P 500 ex-technology earnings growth is meaningfully positive or whether the broader market is increasingly dependent on a small number of mega-cap technology earners to support the index-level growth narrative. Equal-weighted S&P 500 earnings performance compared to cap-weighted performance is the cleanest metric for this analysis, and the Q2 reporting season will produce updated evidence.

The valuation dispersion across sectors means that the marginal investment opportunity depends significantly on which sectors are delivering organic growth versus those that are not. Investors who have been underweight the cyclical and value sectors in favour of mega-cap technology concentration are taking specific bets that the Q2 reporting will either validate or challenge.

Question Three: Is Guidance Language Signaling Capex Moderation?

The forward-looking commentary in Q2 earnings reports — particularly the guidance for capex levels in 2026 H2 and 2027 — is the most informative data about how corporate management actually sees the AI cycle developing. Companies that maintain or increase their capex guidance are signaling continued conviction in the AI revenue thesis. Companies that moderate their capex guidance are signaling more cautious assessment of the AI revenue pace.

The specific commentary to watch includes Microsoft’s capex guidance for fiscal year 2027 (the company’s fiscal year ends in June, so the Q2 calendar reporting will include forward guidance for the new fiscal year), Google’s commentary about Cloud capex sustainability, Meta’s specific framework for AI infrastructure investment, and Amazon’s capex guidance which has been the highest in absolute terms among the hyperscalers.

The signal value of capex guidance changes is asymmetric. Increases in capex guidance are generally positive signals for AI infrastructure investment categories (Nvidia, the broader semiconductor ecosystem, data center REITs, utilities) but neutral-to-mildly-negative for the companies themselves because the increases imply that capex is meeting or exceeding the revenue pace the implicit framework expected. Decreases in capex guidance can be interpreted multiple ways: as moderation reflecting a more sober assessment of AI revenue pace (negative for AI infrastructure beneficiaries, neutral for the hyperscalers themselves) or as improved capital efficiency reflecting better-than-expected operational performance (positive for everyone).

The reading-the-tea-leaves work of distinguishing these scenarios is exactly the kind of qualitative analysis that earnings calls produce. Management commentary about cost discipline, capacity utilisation, and the marginal return on additional capex deployment will reveal whether any capex moderation reflects revenue concerns or operational efficiency.

What Does Not Matter As Much As Headlines Suggest

The headline beat-and-miss statistics on EPS and revenue versus consensus expectations are less informative than the specific underlying questions outlined above. Companies routinely beat consensus by mechanical margins (small beats that reflect guidance management rather than fundamental performance) or miss for specific reasons that do not affect the strategic picture. The market reactions to headline beats and misses often correct themselves within days as more detailed analysis reveals the underlying signals.

The day-of price reactions to individual earnings reports also tend to overweight the immediate beat-and-miss while underweighting the qualitative commentary and forward guidance. The information content of an earnings report is not fully expressed in the price reaction to the headlines; the reaction often gets refined over the following weeks as analysts revise their models based on the more detailed disclosures and management commentary.

The traditional sector relative performance analysis — which sectors are beating and which are missing — is informative at the margin but is itself shaped by analyst expectations that may not have correctly modeled the AI infrastructure cycle, the ex-technology cyclical dynamics, or the various other forces operating on different sectors. Sector dispersion in beat-rate analysis is interesting but should be interpreted carefully rather than used as direct sector rotation signal.

The Specific Companies That Will Reveal Most

The earnings reports that will provide the most information value about the strategic questions outlined above are concentrated in a relatively short list. Microsoft (reporting late July) will reveal Azure AI revenue and capex guidance. Alphabet will reveal Google Cloud growth and Gemini-related revenue signals. Meta will reveal AI-driven advertising revenue improvement and Reality Labs capex commentary. Amazon will reveal AWS growth dynamics, Bedrock commentary, and the broader retail business performance.

Nvidia’s earnings reporting (typically late August for the calendar Q2 fiscal quarter) is the most consequential single report for the entire AI infrastructure thesis. Nvidia’s data center revenue growth, customer concentration commentary, and forward guidance will signal whether the AI compute demand is sustaining at the levels that the hyperscaler capex commitments implied.

TSMC’s Q2 reporting (mid-July) will reveal whether the underlying chip manufacturing demand is sustaining, with implications for the entire semiconductor supply chain. The advanced packaging capacity commentary specifically will signal whether the AI chip supply constraints are easing or persisting.

Outside technology, the major banks reporting in mid-July (JPMorgan, Bank of America, Citi, Wells Fargo) will reveal the credit environment, the net interest margin pressure, and any specific commentary on commercial real estate exposure that affects the regional banking sector. The mega-cap industrials reporting will reveal capex and reshoring trends. The mega-cap consumer companies will reveal consumer spending health.

What This Means for Portfolio Positioning

The Q2 2026 earnings season is unlikely to produce dramatic single-event repositioning across the equity market — the structural questions outlined above are too large to be definitively answered by any single quarter’s reporting. But the season will produce incremental evidence that informs portfolio positioning in specific directions.

If mega-cap AI revenue is pacing with capex, the case for sustained mega-cap technology exposure strengthens, and the AI infrastructure beneficiaries (Nvidia, semiconductors, utilities, data center REITs) continue to support their valuations. If the revenue pace is disappointing, the capex moderation signal that may follow becomes a meaningful headwind for the AI infrastructure supply chain and a relative tailwind for the sectors that have been displaced by AI-focused capital allocation.

If ex-technology corporate America is delivering organic growth, the case for broader market exposure improves, and the equal-weighted index strategies that have lagged the cap-weighted index over the past several years may begin to catch up. If the breadth picture continues to weaken, the concentration risk in the cap-weighted index becomes more acute, and the case for active management that explicitly avoids the concentrated names strengthens.

The Q2 earnings season will produce evidence on all three questions. That evidence will be mixed rather than clean — no single quarter resolves structural debates. The appropriate portfolio response is modest tilts where the evidence is favourable, not dramatic rotations. The more important discipline is reading the reports themselves for the specific data points that matter, not headline coverage that weights beat-and-miss over the underlying structural signals.

What to Ignore and What to Read Carefully When Q2 Numbers Land

William Zinsser’s rule about writing applies to earnings analysis too: strip out everything that is not doing necessary work. Most earnings coverage does not follow this rule. It is full of sentences that restate the headline, repeat what management said on the call, and reach confident conclusions from data points that do not support them. The Q2 2026 earnings season will produce a large volume of this material. Reading carefully through it requires knowing what to ignore.

Ignore the beat rate. In any given quarter, roughly 70-75% of S&P 500 companies beat consensus EPS estimates. This is not because corporate America is consistently exceptional. It is because consensus estimates are set low enough to beat. The beat rate tells you about the relationship between management guidance and analyst estimates. It tells you almost nothing about the underlying health of the businesses reporting.

Ignore revenue surprises that are driven entirely by foreign exchange translation. With the dollar weakening against a broad basket of currencies — partly driven by the BOJ normalization and yen carry trade and structural shifts in reserve allocation — US multinationals with significant international revenue will report better top-line numbers in dollar terms than their local-currency results warrant. This is arithmetic, not business performance. An analyst who leads with the revenue beat without noting the FX tailwind is missing the story.

Read the capex commentary carefully. The question of whether AI is generating returns proportionate to the investment will begin to be answered in Q2 guidance language. Companies that committed to aggressive AI infrastructure buildout in 2024 and 2025 are now far enough into the deployment cycle that investors will start asking — and management will start being held to — concrete revenue attribution. Listen for whether the capex commentary shifts from aspirational to specific. If it remains aspirational after this level of spending, that is informative.

The consumer spending environment is the most important context for reading ex-technology results. US housing market affordability at current mortgage rates means that a significant portion of US households has been effectively locked out of the primary mechanism for consumer balance sheet expansion. Companies with exposure to big-ticket discretionary spending — home improvement, furniture, appliances — will show results that reflect this constraint directly. Companies with exposure to debt-financed consumer spending should show it indirectly, in credit quality metrics if not in top-line revenue.

The autonomous vehicle sector is a useful microcosm for the broader Q2 dynamic: companies where the investment cycle is well ahead of the revenue cycle, where management guidance has repeatedly pushed the monetisation timeline forward, and where the gap between the bull case narrative and the financial results is now large enough that investors are beginning to scrutinise it directly. Every sector in Q2 with a similar structure — large capex commitments, delayed revenue attribution, aspirational guidance — should be read through the same lens.

The Microsoft model of progressive value extraction — building a user base, then progressively monetising it — is the template that most platform businesses are operating from. Q2 is the quarter where investors can begin to assess whether that progression is actually occurring or whether the user base is a leading indicator that has not yet converted to revenue. Companies that can show both metrics moving in the same direction will be differentiated from those where usage is growing and monetisation remains a future promise.

Q2 earnings will tell you something about Q2. The more interesting question is what they tell you about the assumptions embedded in current valuations. Most of those assumptions require answers that Q2 will not yet provide. Reading the results carefully means being honest about what remains unresolved.

Anthropic launched Project Glasswing on April 7, 2026, and gave a restricted group of partners access to Claude Mythos Preview — a model specifically designed to autonomously discover and exploit software vulnerabilities. In the weeks that followed, the model identified more than 10,000 high- and critical-severity vulnerabilities in widely deployed software. It found zero-days in every major operating system. It found them in every major web browser. It autonomously identified and fully exploited a 17-year-old remote code execution flaw in FreeBSD that allowed unauthenticated root access from anywhere on the internet. It found a critical vulnerability in wolfSSL with a CVSS score above 9.1.

Anthropic reported 1,596 verified findings directly to software maintainers. Ninety-seven have been patched. Eighty-eight security advisories have been published.

Claude Mythos remains restricted to approximately 50 vetted partners. It will not be released to the public.

That is the correct decision, and the gap between what the model found and what has been patched is the most important number in this story.

What Claude Mythos Preview Actually Does

Claude Mythos Preview is not a vulnerability scanner in the conventional sense. Traditional vulnerability scanning tools — automated checkers like Nessus, Qualys, or Tenable — identify known vulnerabilities by matching against databases of existing CVEs. They are pattern-matchers. Mythos is a different category of capability.

The model performs autonomous vulnerability research: it reads source code, understands program logic, identifies edge cases in memory management and input handling, generates working proof-of-concept exploits to confirm exploitability, and operates without human guidance on the specific vulnerabilities it pursues. The FreeBSD example is illustrative of this distinction. The RCE vulnerability Mythos identified had existed in the codebase for 17 years. It was not in any CVE database. It had not been identified by any automated scanner or previous security audit. Mythos found it, confirmed it was exploitable, and generated a working exploit that demonstrated full root access from an unauthenticated remote user.

That is not a scanner. That is an autonomous security researcher operating at a scale and speed that no human team can match.

Project Glasswing: The Defensive Framing

Anthropic structured Project Glasswing as a defensive consortium. The access list reads like a who’s who of critical software infrastructure: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. These are the companies whose software and infrastructure, if successfully attacked, would affect hundreds of millions of users and trillions of dollars in financial activity.

The theory of the project is straightforward: if AI can autonomously find zero-days at scale, the question is not whether they will be found — it is whether defenders or attackers find them first. Project Glasswing is an attempt to systematically front-run the attacker cohort by giving defenders priority access to the same capability that offensive actors will eventually develop independently.

CISA’s new 72-hour cyber incident reporting rule, which now covers approximately 300,000 companies across critical infrastructure sectors, becomes substantially more significant in a world where AI-enabled zero-day discovery is available. The reporting mandate assumes that incidents will occur. The Project Glasswing approach attempts to reduce the attack surface before incidents happen — two complementary postures that together describe what enterprise cybersecurity strategy looks like in the AI era.

Anthropic’s stated position, delivered alongside the Glasswing announcement, was blunt: no company currently has sufficient safeguards to defend against the full capability of what Mythos can do. That is not a marketing statement. It is a factual claim about the gap between what AI-enabled offensive capability can achieve and what the current state of enterprise security infrastructure is equipped to handle.

The Patch Rate Problem

The most consequential number in the Project Glasswing results is not the 10,000+ vulnerabilities identified. It is the 97 that have been patched.

Anthropic reported 1,596 verified, high-quality findings to software maintainers. Six weeks later, 97 patches exist. That is a 6% patch rate on confirmed, critical-severity vulnerabilities in widely deployed software. The remaining 94% of reported vulnerabilities remain unpatched in production software running on devices and servers worldwide.

The reason is not negligence by maintainers. The reason is capacity. The open-source software community runs on volunteer maintainers who are already overwhelmed by normal issue volume. Research from the Linux Foundation estimates the average critical vulnerability takes 98 days from discovery to patch deployment across open-source projects. Three thousand concurrent reports from a single disclosure event would stretch that timeline beyond practical resolution. A sudden influx of 1,596 high-quality, confirmed, critical vulnerability reports — each requiring analysis, reproduction, fix development, testing, and coordinated disclosure — represents years of work arriving simultaneously. The maintainers of FreeBSD, wolfSSL, and the dozens of other affected projects do not have the engineering bandwidth to process what Mythos generated in weeks.

This creates a genuinely novel security risk. The vulnerabilities are now known to Anthropic and its 50 Glasswing partners. They are known to the maintainers who received the reports. They are not yet patched. And the same AI capability that identified them is not under Anthropic’s exclusive control for long — adversarial actors will develop comparable capability, either independently or by fine-tuning less safety-constrained models on vulnerability research data.

The window between “vulnerability found by defenders” and “vulnerability patched in the wild” is the attack surface that Project Glasswing inadvertently created by operating faster than the patching infrastructure can process.

Why Restricting Access Is the Correct Decision

The instinctive critique of Anthropic’s decision to restrict Mythos is that it concentrates a powerful capability in a small group of companies — many of them Anthropic’s commercial partners or investors — and withholds it from the broader security research community, which might patch vulnerabilities faster if it had access.

That critique misreads the risk surface. The security research community is not a monolithic defensive actor. It includes researchers who responsibly disclose, researchers who sell findings to governments, researchers who operate in gray markets for zero-day exploits, and actors who are straightforwardly malicious. Releasing Mythos into that environment is not “giving defenders access to a powerful tool.” It is releasing an autonomous exploit generation capability into a population that includes people who will use it offensively.

The zero-day exploit market has functioned for years on the economics of scarcity — a working exploit for a critical vulnerability in a major OS or browser can sell for hundreds of thousands of dollars because finding such vulnerabilities is hard and slow. Mythos makes that economics model obsolete. A public Mythos would collapse zero-day prices not by flooding the defensive community with information, but by flooding the offensive market with exploits that cost nothing to generate.

Anthropic’s broader enterprise strategy has consistently prioritised safety architecture as a competitive differentiator. The Glasswing restriction is consistent with that positioning: the company is making the judgment that the cost of misuse exceeds the benefit of broad access, and it is willing to accept the “concentrating power in large incumbents” critique to maintain that position. That judgment may be right. It may be that the only way to use a capability this dangerous beneficially is to control who has it.

The CVE Infrastructure Stress Test

Project Glasswing is also a stress test of the CVE system itself. The Common Vulnerabilities and Exposures database, managed by MITRE with CISA funding, is the global standard for vulnerability identification and tracking. The CVE assignment process was designed for a world where vulnerability discovery was human-paced — a few hundred high-quality reports per year from the global security research community might generate a few thousand CVEs.

Mythos generated thousands of vulnerabilities in weeks. The CVE numbering authority structure — which relies on CVE Numbering Authorities (CNAs) at individual companies to assign identifiers before coordinated disclosure — is not built for the throughput that AI-enabled discovery can produce. MITRE has been underfunded relative to CVE volume for years; a world where multiple AI systems simultaneously discover vulnerabilities at Mythos-scale throughput would require either a fundamentally restructured CVE process or an acknowledgment that the current system cannot track what is actually being found.

The 88 published advisories from Glasswing represent only the fraction of findings that have proceeded far enough through the disclosure-and-patch pipeline to be public. The 1,596 reported-to-maintainer findings have entered a process that was not designed for this volume, and the output rate — 97 patches in six weeks — suggests the process is already at capacity.

What This Means for Enterprise Security Teams

For enterprise security teams, the Project Glasswing results have two practical implications that operate on different timescales.

In the near term, the findings remind security teams that the patch backlog is not primarily a prioritisation failure — it is a capacity problem. Even with perfect knowledge of critical vulnerabilities, the speed at which patches can be developed, tested, and deployed in enterprise environments is constrained by change management processes, dependency chains, and operational risk tolerance. AI-enabled discovery accelerates the information side of the equation without accelerating the remediation side. The attack surface that exists in the gap is real and growing.

In the medium term, the capability Mythos demonstrates will not remain exclusive to Anthropic’s consortium. Competing AI labs are running comparable research programs. Nation-state cyber programs are almost certainly working on offensive AI vulnerability discovery. The question for enterprise security strategy is not whether AI-enabled zero-day discovery becomes broadly available, but when — and whether the defensive infrastructure and patching capacity exists to respond when it does.

The companies on the Glasswing access list — AWS, Apple, Microsoft, Google — have the engineering resources to process high-volume vulnerability reports and deploy patches at scale. The rest of the enterprise software market does not. The security gap that Glasswing is trying to close is not evenly distributed across the software supply chain, and the portions of the supply chain that are most exposed are not necessarily the ones with Glasswing access.

The Honest Statement Buried in the Announcement

Buried in Anthropic’s project documentation is a statement that deserves to be treated as a headline rather than a footnote: no company currently has sufficient safeguards to defend against what Claude Mythos can do.

That is not Anthropic hedging. It is the company that built the capability acknowledging that the attack surface it can expose exceeds the current state of defensive capacity. The implication is that even the Glasswing consortium members — who include the largest and most sophisticated software security operations in the world — are operating with meaningful unpatched exposure to AI-identified vulnerabilities.

The honest interpretation of Project Glasswing is that Anthropic built something that can find critical vulnerabilities faster than the industry can fix them, is distributing it to a restricted group of defenders to create as much lead time as possible, and is publicly acknowledging that the lead time may not be enough. That is a responsible way to handle a genuinely dangerous capability. It is also a sobering statement about where AI-enabled offensive security capability is relative to the defensive infrastructure meant to contain it.

The Bottom Line

Claude Mythos found 10,000 critical software vulnerabilities in weeks. Six percent of the confirmed findings have been patched. The model remains restricted to 50 vetted partners because releasing it publicly would hand an autonomous zero-day generation capability to a population that includes bad actors.

The patch rate — 97 out of 1,596 reported — is the number that should concern enterprise security teams and policymakers more than the total vulnerability count. It is not evidence that the security community is failing. It is evidence that AI-enabled discovery has outrun the capacity of the human infrastructure meant to respond to it.

Anthropic is right to restrict access. The problem is that restricting access is a delaying action, not a solution. The capability will diffuse regardless of what one company decides. The question that has not been answered — by Anthropic, by CISA, or by the broader security community — is what the infrastructure looks like that can actually process AI-scale vulnerability discovery at the patching speed it requires.

There is a civilizational-scale mismatch embedded in what Project Glasswing has surfaced, and it predates AI. Human institutions have always lagged the capabilities of the technologies they generated — the gap between what a technology can do and what the governance frameworks designed to manage it can actually process has been a consistent feature of every major technical transition, from industrial synthesis to nuclear physics to genetic engineering. What Glasswing changed is the ratio. A single AI model identifying 10,000 critical vulnerabilities in weeks is not a linear acceleration of the existing discovery process. It is a phase transition in the rate at which attack surface expands relative to the institutional capacity built to contain it. The defense layer has not undergone a comparable phase transition. The gap between enterprise AI pilots and production security deployments means that even organizations actively investing in AI-enabled defense are still running their detection, triage, and remediation infrastructure at human-institutional speed, against an offensive capability that now operates at AI speed. Anthropic’s decision to restrict access to Glasswing is the correct response to this asymmetry. The honest observation is that restriction is a delaying action, not a resolution. Closing the gap would require a coordinated rethinking of the regulatory, technical, and institutional infrastructure that processes vulnerability disclosure at the scale AI-enabled discovery now demands — infrastructure that was not designed for the rate at which a model like Glasswing can generate work for it.

On May 14, 2026, the CLARITY Act passed the Senate Banking Committee by a vote of 15-9. This is the first time a comprehensive digital asset market structure bill has cleared a Senate committee. It is a significant milestone for an industry that has operated in regulatory ambiguity for more than a decade — and for policymakers who have long argued that the absence of clear rules has harmed both innovation and investor protection in equal measure.

The vote does not make the CLARITY Act law. It does not even guarantee a full Senate vote. There are substantive obstacles ahead, including a Democratic ethics provision that has become a genuine political sticking point rather than a procedural formality. But the committee clearance establishes the bill as the most advanced piece of comprehensive crypto legislation in US history, and the framework it proposes deserves careful examination for anyone who operates in, invests in, or regulates digital assets.

The Vote: Who Supported It and Who Opposed It

The 15-9 committee vote broke largely along party lines with two notable exceptions. All Republicans on the Senate Banking Committee voted for the bill. Two Democrats crossed the aisle to join them: Senator Ruben Gallego of Arizona and Senator Angela Alsobrooks of Maryland. The other Democratic members of the committee voted against.

The opposition coalition is instructive. Banking industry representatives, major labour unions, and law enforcement agencies all registered opposition to the bill. The banking industry’s concerns centre on competitive displacement — a comprehensive market structure framework for digital assets could enable crypto firms to offer financial services that currently require banking charters, without the full regulatory burden that banks carry. Labour unions have raised concerns about consumer protection provisions they view as insufficient. Law enforcement agencies have expressed concern that the bill’s safe harbour and privacy provisions could complicate illicit finance investigations.

These opposition positions are not simply procedural. They reflect substantive disagreements about whether the bill adequately addresses the real-world harms that have accompanied the growth of digital asset markets: exchange collapses, fraud at scale, and the use of crypto infrastructure for money laundering and sanctions evasion. The sponsors of the bill argue that the illicit finance provisions address these concerns directly. The opponents argue they do not go far enough.

What the CLARITY Act Actually Does

The name is an acronym — the bill’s sponsors were clearly optimising for the brand value of clarity in a regulatory environment that has been anything but. What the bill actually does is establish the first comprehensive statutory framework for how digital assets are classified, regulated, and traded in the United States. Let us go through the major provisions.

Securities vs. commodities classification. This is the foundational problem the bill addresses. Currently, crypto assets exist in a regulatory no-man’s-land. The SEC has argued that most tokens are securities — investment contracts that fall under its jurisdiction. The CFTC has argued that most tokens are commodities — like oil or wheat — that fall under its jurisdiction. The two agencies have overlapping and sometimes conflicting claims, and neither position has been definitively established by statute. The CLARITY Act would establish clear criteria for when a digital asset is a security (SEC jurisdiction) versus a commodity (CFTC jurisdiction), resolving a decade of regulatory confusion that has resulted in regulatory enforcement substituting for regulatory rulemaking. XRP’s regulatory resolution with the SEC is the most prominent example of what enforcement-driven classification looks like in practice — and why the industry wants statutory clarity instead.

Exchange registration. Digital asset exchanges would be required to register with the appropriate regulator based on the classification of the assets they trade. An exchange that lists primarily commodity-classified tokens registers with the CFTC. One that lists primarily security-classified tokens registers with the SEC. This mirrors the existing framework for traditional financial markets, where commodity exchanges register differently from securities exchanges, but applies it to a market structure that did not previously fit neatly into either category.

DeFi framework. Decentralised finance presents the hardest regulatory classification question in the bill. DeFi protocols — smart-contract-based systems for lending, borrowing, trading, and yield generation — do not have identifiable operators in the traditional sense. The protocol runs on a blockchain; the code is the product. The CLARITY Act’s DeFi framework attempts to distinguish between truly decentralised protocols, which would receive lighter-touch regulatory treatment, and centralised entities that describe themselves as DeFi but maintain control over key protocol parameters. The distinction matters enormously for how DeFi projects are required to register, disclose, and operate.

Stablecoin yield limitations. The bill addresses the growing stablecoin yield sector — products that allow holders of dollar-pegged stablecoins to earn interest on their holdings through on-chain lending or other mechanisms. The provisions impose limitations on the yield that stablecoin products can offer, distinguishing between products that function like bank deposits (and should therefore be regulated like deposits) and those that function like investment products (which should carry disclosure requirements). The question of which products fall into which category has direct implications for Tether’s $150B USDT and other yield-bearing stablecoin products currently operating outside this framework. This provision interacts with the GENIUS Act — the stablecoin bill that passed earlier — and the relationship between the two bills’ stablecoin provisions will need to be reconciled.

Tokenisation standards. The bill includes provisions establishing standards for the tokenisation of real-world assets — securities, real estate, commodities — on blockchain infrastructure. Tokenisation has been one of the fastest-growing segments of the crypto market, driven by institutional interest in using blockchain-based settlement to improve the efficiency of traditional asset markets. Clear tokenisation standards would provide the legal certainty that institutional participants need before scaling tokenisation programs.

Developer protections and safe harbours. One of the most politically significant provisions is the safe harbour for developers of decentralised protocols. Currently, software developers who write code that others use to facilitate financial transactions can face regulatory liability based on how that code is subsequently used. The CLARITY Act’s developer protections would shield open-source software developers from regulatory and civil liability for the downstream use of their code, provided certain conditions are met. The legal structure questions that the CLARITY Act’s developer protections address have been central to every DAO and DeFi project’s compliance planning for years.

Customer property protections in bankruptcy. Following the collapse of several major crypto exchanges — most notably FTX in 2022 — the treatment of customer assets in crypto exchange bankruptcies emerged as a critical policy gap. Customers of bankrupt exchanges were treated as unsecured creditors, receiving pennies on the dollar years after the collapse. The CLARITY Act’s customer property provisions would establish that customer assets held by a registered exchange are not the property of the exchange and cannot be used to satisfy exchange creditors in bankruptcy. This provision addresses one of the most concrete consumer harm scenarios that the collapse of FTX made visible.

Illicit finance provisions. The bill includes anti-money laundering and know-your-customer requirements for registered entities, expanded reporting obligations for large transactions, and provisions addressing the use of privacy-enhancing technologies in illicit finance contexts. What regulatory frameworks actually require of exchanges in practice has been tested repeatedly in enforcement actions — the CLARITY Act’s illicit finance provisions attempt to codify those requirements in statute rather than leaving them to be developed through enforcement.

The Ethics Provision: The Real Obstacle

The most significant obstacle to the bill’s passage in the full Senate is not a policy disagreement about crypto market structure. It is an ethics provision that Democratic senators want included, and that Republicans — and the White House — are resisting.

The provision would prohibit government officials from holding, trading, or otherwise financially benefiting from digital assets that they regulate. The Democratic senators supporting this requirement argue that it is a basic conflict-of-interest protection — the same kind of provision that prevents members of Congress from trading stocks in sectors they regulate. Without it, they argue, the bill creates an environment where the officials responsible for crypto regulation have personal financial stakes in the industry’s success, creating an obvious incentive to regulate lightly.

The provision is politically pointed for a specific reason: President Trump and members of his family and administration have publicly known crypto holdings. Trump’s memecoin, launched before and maintained during his presidency, has been a source of ongoing controversy. A prohibition on officials holding crypto assets they regulate would, depending on its scope, require divestitures or recusals that would be politically uncomfortable. Fortune’s coverage of the bill described the ethics provision as “the critical juncture” that could determine whether the bill has sufficient Democratic support to achieve a filibuster-proof majority in the full Senate.

CoinDesk’s reporting confirmed that Democratic senators have been consistent: without the ethics provision, they cannot support the bill on the full Senate floor. The current 15-9 committee vote — which includes only two Democratic votes — is not a sufficient margin for full Senate passage under cloture rules that require 60 votes to overcome a filibuster. The Republican majority alone is 53 votes. To reach 60, the bill needs seven or more Democratic votes in the full Senate. Getting from two to seven requires resolving the ethics provision debate.

The GENIUS Act Distinction

The CLARITY Act is frequently discussed alongside the GENIUS Act, the stablecoin bill that passed earlier in 2026. Understanding the distinction is important for tracking the legislative calendar and the regulatory impact of each bill.

The GENIUS Act dealt specifically with stablecoins — dollar-pegged digital assets used primarily as payment instruments or stores of value within the crypto ecosystem. It established requirements for stablecoin issuers: reserve backing, disclosure, registration, and consumer protection provisions. It did not address the broader question of how non-stablecoin digital assets are classified or regulated.

The CLARITY Act is the companion legislation that addresses everything the GENIUS Act did not. It handles the securities-versus-commodities classification question, the exchange registration framework, the DeFi treatment, and the developer protections that the stablecoin bill explicitly excluded from its scope. Together, the two bills would constitute a comprehensive statutory framework for digital asset markets — the first in US history. Separately, each addresses only a portion of the regulatory question.

The sequencing matters. The GENIUS Act’s passage demonstrated that bipartisan crypto legislation is achievable in the current Senate — a proof of concept that the CLARITY Act’s sponsors have cited in building the case for committee consideration. But the GENIUS Act was a narrower, less contentious bill. The CLARITY Act’s scope is broader, its provisions more complicated, and its political obstacles — including the ethics provision — more significant.

The Path to Law: What Has to Happen Next

Senate committee clearance is step one of a multi-step process. Here is what has to happen for the CLARITY Act to become law.

First, the bill needs to be scheduled for a full Senate floor vote. This is within the control of Senate leadership and the legislative calendar. The White House has set a public target of a July 4 signing. Senator Gillibrand, one of the bill’s key sponsors, has publicly predicted passage in the first week of August. The gap between those two dates reflects the scheduling uncertainty inherent in Senate floor time, where competing legislative priorities — appropriations, nominations, foreign policy matters — can push any individual bill’s floor time.

Second, the bill needs to survive the cloture vote — 60 votes to proceed to a floor vote and end debate. This is where the ethics provision becomes decisive. Senate Democrats who might be inclined to support the bill on the merits will face significant pressure from their caucus to hold firm on the ethics provision as a condition of their vote. If a version of the bill without the ethics provision reaches the floor, the cloture vote may fail.

Third, if the Senate passes a version of the bill, it needs to be reconciled with the House companion bill, HR 3633 in the 119th Congress. The House version has its own provisions that may differ from the Senate version in substantive ways. House-Senate reconciliation on a bill of this complexity typically requires a conference committee or negotiated amendments — another time-consuming process that compresses against the July 4 and August timelines that the bill’s sponsors are targeting.

Fourth, the reconciled bill needs presidential signature. This is the step where the ethics provision has the most leverage. If a conference report includes the ethics provision, the White House may decline to sign it. If it excludes the ethics provision, Democratic senators may withhold the votes needed for cloture. The resolution of this impasse is the defining political challenge for getting the bill across the finish line.

Industry Reaction and the Stakes

The crypto industry has described the CLARITY Act as its top legislative priority for 2026. The committee clearance was met with widespread optimism from major exchanges, investment funds, and protocol developers who have been operating under regulatory ambiguity for years. CoinDesk reported that industry representatives described the vote as a historic milestone — the first time a comprehensive market structure bill had cleared a Senate committee — and expressed confidence that the full Senate vote would follow.

The stakes are significant. In the absence of a statutory framework, crypto regulation in the United States has been conducted primarily through enforcement actions — the SEC and CFTC bringing cases against individual projects and exchanges to establish precedent by litigation rather than rulemaking. This approach has been expensive for the industry and has produced a body of case law that is inconsistent, jurisdiction-specific, and difficult for new market participants to navigate. A statutory framework would replace enforcement-driven regulation with rule-driven regulation — a shift that most market participants across the political spectrum agree would improve regulatory clarity and market confidence.

The international competitive context adds urgency. The European Union’s MiCA framework — Markets in Crypto Assets regulation — came into full effect in 2024 and has established the EU as the first major jurisdiction with comprehensive statutory crypto market structure rules. The UK and several Asian jurisdictions have followed with their own frameworks. The United States, the world’s largest capital market, remains without a statutory framework. This regulatory gap has been cited repeatedly by crypto companies choosing to headquarters outside the US.

Where the Bill Stands: June 2026

The committee vote was May 14. As of early June 2026, the bill has not been scheduled for a full Senate floor vote. The White House’s July 4 signing target requires scheduling, cloture, and passage all within 25 days — a timeline that assumes the ethics provision impasse is resolved quickly and the Senate floor calendar opens. Senator Gillibrand’s August timeline is more realistic given the procedural steps remaining.

The ethics provision remains unresolved. There has been no public indication that Democratic senators have dropped their demand or that the White House has accepted a version of the provision. Until that impasse clears, the cloture math does not work: the current two Democratic committee votes are insufficient to reach 60 in the full Senate, and the senators who have conditioned their support on the ethics provision have not moved publicly.

The industry’s optimism from committee clearance is warranted as a statement of legislative progress. It is less warranted as a prediction about timing. The CLARITY Act has a clearer path to law than at any previous point in crypto regulatory history — but “clearer path” and “imminent passage” are different claims. Market participants should be preparing for a framework that is likely to become law before the end of 2026, while avoiding the operational risk of assuming it will be signed before the summer recess.

What Committee Clearance Actually Means

Committee clearance is not passage. The CLARITY Act has cleared the Senate Banking Committee; it has not passed the full Senate, been reconciled with the House version, or received presidential signature. The path from committee clearance to law is real and navigable, but it requires resolving the ethics provision impasse, finding floor time in a compressed legislative calendar, and navigating House-Senate reconciliation on a complex bill.

What committee clearance does establish is this: the CLARITY Act has survived detailed legislative scrutiny. It has been marked up, amended, and debated in the committee with primary jurisdiction over financial regulation. It has bipartisan support — narrow, but real. It has the backing of the White House and Senate Republican leadership. It has a companion House bill. The legislative infrastructure for passage exists.

Whether the political will to resolve the ethics provision impasse materialises before the summer recess will determine whether 2026 becomes the year the United States finally established comprehensive rules for digital asset markets — or whether the industry enters 2027 still waiting for the statutory clarity it has pursued for a decade.

Sources

• CoinDesk — “Clarity Act clears U.S. Senate committee, on its way to a final test in Congress” (May 14, 2026)
• CNBC — “Crypto industry scores win as Clarity Act regulation bill clears Senate hurdle” (May 14, 2026)
• CoinDesk — “Clarity Act, in the flesh, unveiled by U.S. Senate Banking Committee” (May 11, 2026)
• Fortune — “The crypto industry’s Clarity Act hits a critical juncture” (May 13, 2026)
• Davis Wright Tremaine — “Senate Banking Committee Advances Crypto Market Structure Bill”