Author: Dan Santarina

On April 27, 2026, Microsoft and OpenAI announced a restructured partnership. The headline detail — which received less attention than it deserved — was that the exclusivity arrangement at the centre of the original deal had been ended. OpenAI could now offer its models on AWS, Google Cloud, Oracle, or any infrastructure it chose. On April 28, GPT-5.5, Codex, and OpenAI’s managed agents appeared on Amazon Bedrock. The exclusivity that Copilot was built on lasted until the Monday press release. By Tuesday, Amazon was selling the same models.

Microsoft invested more than $13 billion in OpenAI across three tranches. The investment was financial but its strategic value was structural: it gave Microsoft an exclusive right to deploy OpenAI’s models through Azure and an exclusive claim on OpenAI’s compute infrastructure for training and inference. Enterprise buyers who wanted GPT-4 or GPT-4o in production faced a practical requirement to use Azure. That created a dependency relationship that benefited Microsoft’s cloud business, provided the model quality foundation for Copilot, and placed Microsoft two or three years ahead of competitors in enterprise AI deployment.

That structural advantage has been restructured into a competitive market. The implications for Copilot are more significant than Microsoft’s public communications have acknowledged.

What the Exclusivity Was Worth

The original Microsoft-OpenAI arrangement, structured across investments beginning in 2019 and significantly expanded in 2023, gave Microsoft three distinct advantages in enterprise AI.

The first was infrastructure exclusivity. OpenAI’s training runs, inference workloads, and customer-facing API traffic ran through Azure. This made Microsoft the backbone of the world’s leading AI lab, generating significant Azure revenue and creating deep engineering integration between OpenAI’s research teams and Microsoft’s cloud infrastructure. Competitors could build their own AI capabilities, but they could not build products on OpenAI’s models through their own infrastructure. The compute exclusivity was a moat measured in billions of dollars and years of specialised engineering.

The second was model exclusivity. Enterprise buyers who wanted to deploy GPT-4, GPT-4o, or the O-series reasoning models in production workloads had to do so through Azure OpenAI Service. AWS, Google Cloud, and Oracle could not offer these models. This meant that enterprise procurement conversations about frontier AI model deployment converged, functionally, on a binary choice: Microsoft or build your own. For most large organisations, “build your own” was not a realistic option. Azure OpenAI Service was the enterprise frontier model market.

The third advantage was product integration. Microsoft embedded GPT capabilities across its productivity suite — Word, Excel, PowerPoint, Teams, Outlook, and the broader Microsoft 365 ecosystem — through Copilot. The integration was deep enough that Microsoft could argue, credibly, that using OpenAI models inside the applications where enterprise employees already worked was categorically different from using ChatGPT in a browser tab. The model quality plus the integration context was the Copilot value proposition.

The April 27 restructuring removed the first two advantages. What remains is the third.

Why OpenAI Ended It

The restructuring did not happen because Microsoft wanted it to. The terms were renegotiated because OpenAI’s market expansion had been structurally limited by the exclusivity arrangement.

OpenAI’s revenue chief Denise Dresser, in an internal memo cited publicly after the April 27 announcement, stated that the exclusive Azure arrangement had “limited our ability to meet enterprises where they are.” The practical problem: enterprise AI procurement decisions are made inside existing cloud commitments. AWS holds the largest enterprise cloud market share. Google Cloud has deep penetration in data infrastructure and analytics. Many enterprise technology leaders who wanted to evaluate OpenAI’s models faced an implicit requirement to expand their Azure footprint — a procurement decision that required additional budget approval, contract renegotiation, and multi-quarter implementation cycles.

OpenAI was losing deals not because its models were inferior but because the distribution channel was constrained. Exclusivity was the price OpenAI paid for Microsoft’s investment. When the cost of that price — in lost enterprise customers — exceeded the benefit of the Microsoft relationship, OpenAI negotiated the terms.

The restructuring reflects a shift in power within the partnership. In 2019 and 2023, OpenAI needed Microsoft’s capital to continue operating and scaling. By 2026, OpenAI had achieved a valuation north of $300 billion, a revenue run rate that made it self-sustaining, and a model portfolio — GPT-4o, O3, GPT-5.5 — that gave it genuine negotiating leverage. The partner that once needed Microsoft’s balance sheet now had the leverage to renegotiate the terms that constrained its market reach.

The restructuring is detailed in the announcement, but the strategic implication requires unpacking: OpenAI’s decision to prioritise its own distribution over Microsoft’s exclusivity is a statement about where OpenAI believes its long-term value lies. It lies in model quality and developer ecosystem, not in Azure dependency. That is a legitimate and arguably correct assessment. It is also, from Microsoft’s perspective, a significant departure from the terms that justified the investment.

What Microsoft Retained

The restructuring was not total. Microsoft retained meaningful advantages, and it is worth being precise about what they are.

The license to deploy OpenAI models on Azure continues and has been extended through 2032. Microsoft can continue to offer Azure OpenAI Service to enterprise customers, and the existing enterprise deployments built on that service do not need to be migrated. For the installed base of Azure OpenAI customers — substantial, given that the product has been available since 2023 — the restructuring changes the competitive landscape but does not disrupt their current deployments.

Microsoft also retained a four-month first-mover window on new frontier model releases. Any new OpenAI model will debut on Azure before it becomes available on competing cloud platforms. For the specific use case of enterprises wanting access to the most capable OpenAI models at launch, Azure remains the first option. GPT-5.5 appeared on AWS Bedrock on April 28 because it had already been available on Azure before the restructuring announcement. Future models will follow a four-month Azure-exclusive window before broader distribution.

Microsoft also retained its integration depth. The Work IQ API, generally available from June 16, gives Copilot agents access to Microsoft 365 data — calendar patterns, document activity, communication signals — in ways that no competing platform can replicate through the Bedrock integration alone. An enterprise deploying GPT-5.5 through AWS Bedrock gets a capable model. An enterprise deploying Copilot with Work IQ gets a capable model grounded in the organisation’s own operational data.

These are real advantages. They are also narrower than what existed before April 27.

What Copilot Now Faces

The competitive landscape for enterprise AI has changed materially since April 28. An enterprise technology leader evaluating AI assistant options in June 2026 can now access GPT-5.5 and Codex through Amazon Bedrock without an Azure commitment. They can deploy OpenAI-powered agents through AWS infrastructure, benefit from Amazon’s enterprise support and pricing flexibility, and maintain their existing AWS relationship rather than expanding into a second cloud provider.

This matters because the enterprise objection to Copilot was never primarily about model quality. The model quality argument — GPT is better than Google’s models, therefore Copilot is better than Google Workspace AI — was always partially correct but rarely the deciding procurement factor. Enterprise AI adoption decisions involve vendor concentration risk, cloud spending commitments, data sovereignty requirements, integration complexity, and total cost of ownership across multi-year contracts. An enterprise that had resisted Azure expansion because of those factors now has a path to frontier OpenAI models without that expansion.

Copilot’s remaining competitive argument is integration depth — the M365 context that Work IQ enables, the calendar and document signals that ground agents in operational reality rather than general capability. That argument is correct and likely valuable to enterprises already running significant Microsoft 365 workloads. It is less compelling to enterprises evaluating a greenfield AI deployment, or to the substantial share of enterprises running hybrid environments where Microsoft 365 is one of several productivity platforms.

The product itself has not helped this argument. Copilot achieved 3.3% paid penetration of Microsoft’s addressable enterprise base with the exclusive model advantage in place. Its accuracy Net Promoter Score deteriorated to -24.1 in September 2025 before partially recovering to -19.8 in January 2026. Among users who have lapsed — who provisioned Copilot and stopped using it — 44.2% cite distrust of answers as the primary reason. The adoption failure predates the exclusivity removal; the product was underperforming before Amazon was given access to the same underlying models.

The exclusivity removal does not explain Copilot’s existing adoption problem. What it does is remove the structural protection that limited how directly that problem could be competed against. The competitor you face when you have exclusive model access and the competitor you face when the same model is available on AWS Bedrock are categorically different in their leverage.

The MAI Concession and the Timeline Problem

At Build 2026 in early June, Microsoft announced the MAI family of seven in-house AI models, developed internally and designed to reduce dependency on OpenAI’s external models over time. The MAI announcement was noted in our coverage of the agentic pivot as an implicit concession on OpenAI dependency. The April 27 restructuring makes that concession explicit: Microsoft is now building an alternative because it can no longer rely on exclusive access to the best available external models.

The MAI family is a strategically correct response. Every major technology company with significant AI exposure — Google with Gemini, Amazon with Nova and Titan, Meta with Llama, Apple with Apple Intelligence — has invested in proprietary model capability rather than depending entirely on external providers. The strategic case for in-house models is real: control over capability roadmap, independence from partnership terms, ability to fine-tune for specific use cases without the constraints of third-party agreements.

The timeline problem is also real. MAI is currently deployed in two Azure data centres. GPT-5.5 is available across AWS Bedrock’s global infrastructure from day one of the non-exclusive arrangement. Google’s TPU 8i delivers approximately 80% better performance per dollar than comparable infrastructure according to Alphabet’s own benchmarks. Amazon’s custom silicon — the Trainium and Inferentia families — now supports a $20 billion annual run rate in AWS AI services. Microsoft’s in-house model ambitions are credible and well-resourced. They are also years behind the infrastructure and model capability of the competitors they are designed to offset.

The four-month first-mover window on new OpenAI frontier models is the bridge between the dependency that is ending and the in-house capability that is being built. It is a meaningful bridge for the specific segment of enterprise buyers who prioritise frontier model access above all else and who are willing to accept Azure as the deployment platform to get it. It is a less meaningful bridge for the broader enterprise market that the restructuring has now opened to competing cloud providers.

The Capex Paradox

Microsoft’s AI capex commitment — $30.88 billion in the most recent quarterly run rate, against a $190 billion annual target — is now simultaneously funding two things that point in different directions.

A portion of that capex funds Azure infrastructure that serves OpenAI’s non-exclusive API customers. When an enterprise deploys GPT-5.5 through AWS Bedrock, some of the inference traffic ultimately runs on infrastructure that may include OpenAI’s compute capacity — which is no longer exclusively Azure. Microsoft is building and maintaining infrastructure that partially serves OpenAI’s expansion to Microsoft’s cloud competitors. The exclusivity removal means that some of the Azure capex investment is now supporting the competitive case for AWS as an AI platform, via OpenAI’s multi-cloud availability.

Another portion of that capex funds MAI development and deployment — the in-house alternative to the external model dependency that the restructuring has made more complicated. Microsoft is spending to build what it used to have exclusive access to. The $13 billion investment bought a period of exclusive advantage and, when that advantage was restructured away, a residual license and a four-month window. The MAI investment is the attempt to build a durable replacement for what the original investment provided.

The sum of those two expenditures — maintaining Azure infrastructure that partially benefits competitors and funding in-house model development against a multi-year timeline — is the financial structure of Microsoft’s AI position after the restructuring. The peer comparison has already established that Microsoft is the only major AI-spending company being punished by the market for its capex allocation. The April 27 restructuring adds a new dimension to that question: the capex is now funding both the dependency Microsoft is trying to escape and the alternative it is trying to build.

What This Means for the Copilot Argument

Copilot’s market positioning has rested on a three-part argument: the model is the best available (GPT), the model is uniquely accessible through Microsoft’s platform (exclusive), and the model is grounded in the enterprise’s own data (M365 integration). The first two parts of that argument were interdependent. “Best model, exclusively available here” is a compelling enterprise value proposition. “Best model, also available on AWS and Google Cloud, but with better Microsoft 365 integration here” is a correct but significantly weaker one.

The integration depth argument — the third pillar — is where Microsoft’s residual competitive advantage now sits. Work IQ API, which provides Copilot agents with access to calendar patterns, document activity, and communication signals from Microsoft 365, offers something that cannot be replicated through a Bedrock integration alone. An enterprise that has committed heavily to Microsoft 365 and wants AI agents grounded in its operational data has a genuine reason to deploy Copilot rather than a Bedrock-hosted GPT alternative. The contextual intelligence that comes from deep M365 integration is real and valuable.

The problem is the prerequisite. The integration argument is most compelling for enterprises that have already committed to Microsoft 365 as their primary productivity platform, are deploying AI at sufficient scale to benefit from the contextual grounding Work IQ provides, and have IT infrastructure capable of implementing Copilot’s agent architecture at the reliability levels that autonomous agents require. The over-extractive incumbency dynamic that Microsoft faces — where aggressive monetisation of the M365 install base creates resistance to additional AI licensing spend — means the integration argument is not automatically persuasive even to the enterprise customer most likely to benefit from it.

The agentic pivot announced at Build 2026 adds a further dimension. Microsoft has repositioned Copilot from a productivity assistant to an autonomous agent platform — a shift that, as we have documented, requires categorically higher reliability than the sidebar assistant use case. The reliability record — the June 1 outage that affected 14,000 users three days before the Build keynote — does not yet support the autonomous agent ambition. The model that supports that ambition is now also available on AWS Bedrock, where enterprise buyers can evaluate it against Microsoft’s platform before deciding where to deploy.

The Investment Return Question

Microsoft’s $13 billion investment in OpenAI generated a period of exclusive model access that ran from 2023 through April 2026 — approximately three years. During that period, Microsoft built Azure into the leading enterprise AI cloud, established Copilot as the most prominent AI productivity assistant, and positioned itself as the default enterprise AI partner for the Fortune 500. The exclusivity was the structural foundation for all three.

Whether that foundation justified the investment depends on what was built on top of it. The Azure AI business is substantial: $37 billion in annual AI revenue run rate, 40% year-over-year growth. The Azure position was established during the exclusivity period and does not disappear with the end of exclusivity — Azure OpenAI Service customers have committed infrastructure, built integrations, and trained teams. Installed base in enterprise technology has genuine switching costs.

The Copilot business is more complicated to assess. The product that was supposed to be the consumer of the Azure AI infrastructure and the primary monetisation vehicle for Microsoft’s OpenAI investment achieved 3.3% paid penetration of the addressable enterprise base over two years of exclusivity. The Nadella-level internal designation — Code Red, first applied to Copilot’s enterprise adoption in early 2026 — is the internal acknowledgment that the exclusivity period did not produce the commercial outcome the investment required.

The question now is whether the integration depth, the four-month window, and the MAI roadmap can produce an outcome over the next three years that the exclusivity, the model quality, and the first-mover advantage did not produce in the preceding three. That is not a rhetorical question. It is the actual business question that Microsoft’s AI leadership team is working to answer. The answer will determine whether the $13 billion investment, and the much larger capex commitments that followed it, will produce a return commensurate with the scale of the bet.

The April 28 Bedrock availability did not make Microsoft’s AI position nonviable. Azure’s installed base, the Work IQ integration advantage, the four-month model window, and the MAI roadmap represent a real and well-resourced competitive position. They represent a different competitive position than the one Microsoft occupied on April 26.

On April 26, the enterprise buyer who wanted frontier AI models in production had one cloud option. On April 28, they had at least two, and soon three. The market that was a practical Microsoft monopoly during the exclusivity period became, in the space of a press release and an overnight AWS integration, a competitive market. Competitive markets produce different dynamics than monopoly markets — different pricing pressure, different customer leverage, different switching cost calculus.

Copilot will compete in that market carrying a penetration rate of 3.3%, an accuracy NPS of -19.8, and a product that the company’s own leadership internally designated as a Code Red case two years into its commercial launch. The competitive argument that remains — integration depth, contextual grounding, M365 ecosystem coherence — is the right argument for a specific and substantial segment of the enterprise market. It is a narrower argument than the one that the exclusivity period made available.

Microsoft knew this was coming. The MAI announcement, the Work IQ API, the agentic pivot — each reflects preparation for a post-exclusivity competitive environment. The preparation was real. So was the three-year window that the exclusivity provided, during which Copilot needed to establish the kind of enterprise penetration that would make the transition less exposed. That window closed in April. The penetration rate, at 3.3%, is the record of what was built during it.

On June 8, 2026, Sam Bankman-Fried filed a formal presidential pardon application with the United States Department of Justice Office of the Pardon Attorney. He is 34 years old. He is serving a 25-year sentence at the Federal Correctional Complex in Terre Haute, Indiana, for what a federal jury determined was the largest fraud in cryptocurrency history — the deliberate theft of approximately $8 billion in customer deposits from the FTX exchange. The filing made global news within two hours. FTT, the native token of his bankrupt exchange, jumped 50% before the day was out.

Understanding what this filing actually is — and what it reveals — requires reading it precisely. SBF did not request early release. He did not request a sentence reduction. He did not request a commutation. He requested a “pardon after completion of sentence” — a specific designation that, if granted, would restore certain civil rights once his full term ends. He would remain incarcerated until approximately 2049. The pardon, if Trump approved it tomorrow, would not move his release date by a single day.

The narrowness of the request is revealing. A man seeking to escape prison would petition for clemency or a commutation. A man who has concluded that his conviction is a negotiating position, and that the correct negotiating partner is the sitting president, petitions for the symbolic restoration of rights that will not vest for twenty-three years.

This is not a legal argument. It is a theory of accountability.

The Fraud, Established

FTX launched in 2019 and grew, on the basis of aggressive marketing, celebrity endorsements, and Bankman-Fried’s projection of professional credibility, to a peak valuation of $32 billion. At its height, FTX was the second-largest cryptocurrency exchange in the world by volume. SBF was its public face: the effective altruism devotee who slept on a beanbag, the philanthropist who donated to pandemic preparedness and political causes, the founder who testified before Congress about the need for sensible crypto regulation.

What the exchange’s customers did not know, and what the trial established, was that FTX customer deposits were being routed to Alameda Research — Bankman-Fried’s affiliated trading firm — and used as working capital for proprietary trading, political donations, venture investments, and personal expenditure. There was no segregation of funds. Customer balances shown on the FTX platform did not correspond to assets held in custody. The exchange was operating as a fractional reserve, without disclosure, and without the reserve.

When cryptocurrency markets declined in late 2022 and Alameda’s positions deteriorated, the gap between customer balances and actual assets became impossible to conceal. A CoinDesk report on Alameda’s balance sheet in November 2022 triggered a bank run. FTX froze withdrawals within days. The exchange filed for Chapter 11 bankruptcy on November 11, 2022. Eight billion dollars in customer funds could not be returned because they were not there.

One million customers — retail traders, small investors, people who had been told, explicitly and repeatedly, that their funds were safe — discovered their deposits were gone.

The federal trial was thorough. Former FTX executives testified against Bankman-Fried under cooperation agreements. Caroline Ellison, who ran Alameda Research, testified that Bankman-Fried had directed the commingling of customer and trading funds and had been aware of the gap. The jury deliberated for fewer than five hours. The verdict was guilty on seven counts. Federal Judge Lewis Kaplan, at sentencing, stated that Bankman-Fried “knew what he was doing was wrong.” The sentence was 25 years.

The victims’ tally, established by court findings: $8 billion in customer deposits lost; $1.7 billion in investor losses; $1.3 billion in losses to lenders to Alameda Research. Trump, asked about a pardon for Bankman-Fried in January 2026, cited “the scale of the $11 billion fraud” as the reason he had no intention of extending clemency. He restated that position when asked again.

From Prison: The Rewrite

Accountability contestation in cases like this follows a recognizable sequence. The first phase is denial: the events that occurred were not what they appeared to be. The second phase is reframing: the harm that resulted was caused by factors beyond the founder’s control. The third phase is grievance: the people tasked with managing the aftermath made it worse. SBF has worked through all three phases with more speed and visibility than is typical, communicating through prison-approved channels and intermediaries in a manner that has generated a steady stream of copy since his sentencing.

The denial: FTX was “never technically insolvent,” Bankman-Fried has argued through statements from prison. His theory holds that if the exchange had been permitted to restructure — rather than filing for Chapter 11 — customers could have been made whole. The implication is that the harm was a function of the process, not of the $8 billion gap that preceded it.

The reframing: cryptocurrency prices have recovered significantly since the 2022 collapse. Customers whose dollar deposits are being repaid through the bankruptcy proceedings have, in some cases, received amounts that nominally exceed their original balances. SBF has cited this as evidence of his original thesis — that FTX was essentially sound and that the bankruptcy was unnecessary. The framing excludes, entirely, the opportunity cost: customers who held Bitcoin and Ethereum through a four-year bankruptcy freeze missed one of crypto’s most substantial recovery periods, without access to their assets, without the ability to manage their positions, and without the ability to make other investment decisions with frozen funds.

The grievance: the bankruptcy professionals — Sullivan & Cromwell and associated restructuring advisors — have charged more than $1 billion in professional fees managing the FTX estate. SBF has argued, through his representatives, that these costs accelerated and amplified customer harm. The complaint misses the causal structure: the billion-dollar fee is a consequence of the $8 billion fraud, not a cause of it. You do not incur $1 billion in restructuring fees on a company that managed its customer assets appropriately.

None of these reframings is a legal argument for a pardon. None of them engage with what the court actually established. They are, collectively, an ongoing attempt to install an alternative narrative in the space between “convicted of fraud” and “the fraud itself.”

The Political Alignment Campaign

In early March 2026, a post appeared on X attributed to Sam Bankman-Fried’s account, written through prison-approved communications and routed through intermediaries. It praised President Trump’s decision to launch military strikes against Iran as “the right call” and framed the action in national security terms that tracked closely with the administration’s public messaging. The post was notable for what it was not: a statement about cryptocurrency markets, the FTX case, or any subject Bankman-Fried might be reasonably expected to have views on.

Further posts followed. SBF argued that Trump had “saved the Securities and Exchange Commission” by replacing former chair Gary Gensler with Paul Atkins — a crypto-industry-friendly regulatory appointment that the broader digital assets community welcomed. He highlighted lower gasoline prices during the Trump administration. He referenced the administration’s executive orders on digital assets favourably. He expressed support, implicitly and explicitly, for the framework of pro-digital assets policy that the administration has pursued.

The positions, taken individually, are unremarkable. Plenty of people praised the Iran strikes. Many crypto founders welcomed the Atkins appointment. What makes the pattern notable is its context: these statements are not the political views of a free citizen engaging with current events. They are communications from a man serving 25 years for fraud, addressed to an audience of one, through a medium available to him precisely because his communications are monitored and constrained. Every statement requires a decision about what to say through a limited and audited channel. SBF chose, repeatedly, to say things that aligned with the president’s stated positions.

The calculation is not subtle. Trump has pardon authority. Trump has used it, in his second term, for individuals whose cases generated political interest. The pro-crypto regulatory environment suggests some sympathy for digital assets broadly. The path from Terre Haute to a post-sentence pardon, SBF’s apparent theory holds, runs through visible and on-record alignment with the political priorities of the person who holds that authority.

There is something almost formally legible about this. It is the same structured cost-benefit analysis — applied to political capital rather than financial leverage — that characterised Alameda’s operation. Identify the lever. Apply calibrated pressure. Model the expected output.

The problem, in both cases, is that the models assume the rules of the system apply uniformly and that the outcomes can be engineered through the correct inputs. In both cases, that assumption may be wrong.

The Filing Itself

The June 8 application was submitted through the standard channel: the Office of the Pardon Attorney, within the Department of Justice, which processes pardon applications from convicted individuals and forwards recommendations to the White House. The office receives hundreds of applications per year. The process is opaque and there is no statutory timeline for review.

The specific relief requested — “pardon after completion of sentence” — is a precise designation in pardon law. It does not reduce the sentence. It does not trigger any early release mechanism. If granted by the president and certified by the Attorney General, it would take effect when Bankman-Fried’s sentence ends, whenever that is. The civil rights that would be restored include, primarily, the right to vote, the right to hold federal office, and the removal of certain restrictions that federal felony convictions impose on professional and civic participation.

The filing is, in legal terms, a narrow and technically proper request. SBF is entitled to apply. The Pardon Attorney is obliged to process the application. Nothing about the submission is irregular. What makes it worth examining is not its legal form but its strategic function: the filing converts the private alignment campaign — the Iran tweets, the SEC commentary, the gas price observations — into a formal, on-record request that the president’s team must acknowledge and respond to.

Trump’s response, through the White House, was to point to his January 2026 New York Times statement that he had “no intention of pardoning” Bankman-Fried, and to confirm that position remained unchanged. That is the second on-record presidential rejection of an SBF pardon in 2026.

The Polymarket prediction market placed the probability of a pardon by year-end at 8% following the news.

The Contradiction at the Centre

There is a structural problem at the heart of SBF’s position that has received less attention than it deserves. In public statements from prison — through his X account, through interviews, through communications relayed by intermediaries — Bankman-Fried has maintained that he did not steal customer funds. In his Fox Business prison interview, he stated this explicitly: he had not stolen user funds, the bankruptcy process manufactured the crisis, customers were being made whole by price recovery.

That is a claim of innocence. It is also incompatible with the pardon request.

A presidential pardon is not an exoneration. It does not vacate a verdict. It does not establish that a conviction was wrong. It is an act of executive clemency that acknowledges a criminal conviction and extends forgiveness for it. The legal consequence of a pardon is the removal of certain civil penalties; the legal effect on the underlying verdict is nil. A pardoned person remains convicted of the crimes of which they were found guilty.

If SBF did nothing wrong — if FTX was never technically insolvent, if the harm was caused by the bankruptcy professionals, if the trial was a miscarriage — then the correct legal avenue is an appeal arguing that the verdict was unsound. SBF has pursued appeals, without success. He is entitled to continue that path. But an appeal says “the conviction was wrong.” A pardon request says “please forgive the conviction.” Filing both simultaneously requires holding two positions that cannot both be true.

The contradiction is informative. It suggests that what SBF is actually doing is not constructing a coherent legal argument but managing multiple audiences simultaneously: telling supporters he was wrongly convicted, telling the president he deserves mercy, and positioning for whichever avenue produces a better outcome. This is recognizable behaviour in accountability contestation. The frame shifts to match the audience. The goal is not a settled account of what happened but a better position in the ongoing negotiation over what it means.

FTT: The Accountability Market

FTT — the FTX exchange token — is, by any functional measure, a dead asset. The exchange that gave it utility collapsed in November 2022. There is no active development team for FTT. There is no roadmap. There is no product, no fee discount mechanism, no staking yield, no redemption right. The token exists as an entry on a blockchain ledger that used to correspond to something and no longer does. Its all-time high was approximately $85 per token in 2021. In early June 2026, it was trading at $0.21.

On June 8, 2026, within hours of the pardon application news breaking, FTT jumped 50% — rising from $0.21 to $0.35 before pulling back. Trading volume surged. The move was coordinated with the Polymarket estimate of an 8% pardon probability by year-end.

What are traders buying when they buy FTT at $0.35? There is no income case. There is no utility case. The only recoverable case for FTT is something like: SBF is pardoned, the political environment shifts dramatically enough for an FTX reconstitution or successor entity, and FTT acquires some future value in that reconstituted structure. The probability of that chain of events is very low. The 8% Polymarket figure covers only the pardon. The chain beyond it — exchange reconstitution, FTT utility restoration — would require further steps, each with their own probability.

The 50% intraday move on 8% probability, in a very illiquid market, is mathematically coherent. What it represents substantively is a liquid, public, continuously updating market for the probability that a fraud conviction can be politically reversed. That market exists. It is priced and tradeable. It moved on news of a pardon application.

This is not a comment on the traders. Markets price what information is available, and a pardon filing is information. It is a comment on the structure of accountability in the space. If the consequences of building a fraudulent exchange are contingent on the political preferences of a president — if that contingency is liquid and speculative — then the founding condition of the accountability system is weakened. Fraud carries a 25-year sentence unless political proximity generates a discount. The discount is now priced.

The Pattern

SBF’s pardon request is the most visible instance of a pattern that has characterised the crypto founder accountability record across multiple cases in the past eighteen months.

When Bitcoin Depot filed for bankruptcy in May 2026 — collapsing from a $1.6 billion SPAC peak valuation to an $8.9 million enterprise value, with revenue down 49% in a single quarter — CEO Alex Holmes attributed the failure to “increasingly stringent state regulations” and a “regulatory landscape becoming markedly unfriendly.” The regulatory explanation was the dominant frame in the company’s public communications. It was also directly contradicted by the company’s own record: the Attorneys General of Massachusetts and Iowa had sued Bitcoin Depot in February 2025 — fourteen months before the bankruptcy filing — for facilitating over $20 million in cryptocurrency scams targeting elderly residents. The regulators were not acting ahead of Bitcoin Depot’s problems. They were responding to documented harm that the company had failed to address. The accountability frame chose blame over record.

When Christopher Delgado was arrested in February 2026 on charges of running the $328 million Goliath Ventures Ponzi scheme — having promised investors guaranteed 3–8% monthly returns from cryptocurrency liquidity pools while placing less than 0.5% of their money into any pool at all — his public statements included expressions of remorse. “I failed them,” in some form, was the accountability frame offered. The problem: Delgado had, at the time of his arrest, spent investor funds on a James Bond-themed holiday party at the Fontainebleau Miami Beach, six homes in Central Florida including an $8.5 million Isleworth mansion, private jets, Lamborghinis, and what investigators described as an extravagant personal lifestyle maintained in apparent awareness that a federal investigation was underway. The remorse was offered at a significant geographic and temporal distance from the harm done.

The spectrum — regulatory blame, theatrical remorse, political positioning — covers different modes of accountability contestation. Each operates on the same foundational premise: that the accountability outcome is not settled, that it is subject to revision through the correct framing, and that the harm done to specific individuals can be refracted through a narrative that places the founder at a sufficient remove from direct responsibility.

The amateur leadership pattern in crypto is not primarily about technical incompetence — though that is present — or about inexperience with scale — though that too applies. It is about a relationship with accountability that treats consequences as provisional, outcomes as negotiable, and the harm done to users as a context for the founder’s narrative rather than its subject.

SBF’s pardon request is simply the most formally naked version of that relationship. He has taken it to its logical endpoint: a document filed with the Department of Justice, requesting that the president of the United States formally revise the accountability outcome on political grounds.

What the Strategy Reveals

The pardon strategy can be reconstructed from the evidence: the Iran tweets, the SEC commentary, the gas prices, the formal filing. The theory underneath it is legible.

Conviction, in SBF’s apparent view, is not a settled determination of fact. It is a legal outcome that exists in a broader political context. Presidential pardons are real. They have been exercised in the current administration. The digital assets community has significant political standing at the moment. Political alignment — visible, on-record, addressing the president’s specific policy positions — creates or sustains a non-zero probability of clemency. The correct response to that probability is to invest in it by maximising the alignment signal.

This is, structurally, the same analysis that ran Alameda Research. Identify the variable that can be moved. Model the output. Apply calibrated pressure. The fraud operated on the premise that customer fund segregation rules, fiduciary obligations, and fraud statutes were constraints that could be navigated through sufficiently sophisticated positioning. The pardon strategy operates on the premise that a jury verdict and a 25-year sentence are constraints that can be navigated through sufficiently sophisticated political positioning.

The question is whether the second analysis is more accurate than the first. The first one was wrong — the constraints were real, the enforcement was real, the jury was not moved by the sophistication of the positioning. The second one has two data points so far: Trump said no in January, and said no again in June.

The Polymarket traders at 8% are pricing a third data point that is possible but not yet evident. The FTT market at $0.35 is pricing the full speculative chain beyond that. These are reasonable markets to make. They are also, in aggregate, a continuous public statement that the accountability outcome for the largest fraud in crypto history is being treated as a function of political proximity, not as a function of the facts that the jury found.

That treatment is not unique to crypto. Presidential pardons have always been political. What is different here is the scale and visibility of the accountability contestation that preceded the filing: the revisionist claims from prison, the innocence assertions, the simultaneous pursuit of appeal and pardon, the political alignment campaign, and the speculative market for the outcome. The machinery for contesting accountability is more developed, more liquid, and more publicly legible than it has been in previous financial fraud cases.

Whether that machinery will produce a different outcome for SBF than the jury verdict produced is the live question. The current evidence — two presidential rejections, 8% Polymarket odds — does not suggest it will. But the machinery exists, it is running, and its existence is itself a data point about what accountability means in this space.

Trump has said no twice. The Office of the Pardon Attorney will process the application in due course and forward a recommendation, which may or may not be followed. Polymarket will continue updating the probability. FTT will trade at whatever price the speculation supports at any given hour.

Somewhere in Terre Haute, one million people’s former counterparty is working through the calculation that brought him to this filing: that the correct response to a 25-year fraud sentence is to identify the political lever, align visibly with the person holding it, and wait for the probability to move.

What those one million people received, in the interim, is not an acknowledgment that the fraud was a fraud. What they have received is a claim that it was never technically insolvent, a complaint about the fees charged to manage the wreckage, a series of political endorsements from a federal prison cell, and now a formal request for a pardon framed as the restoration of civil rights.

The pardon request asks for something narrow — voting rights and professional freedoms that vest in 2049. What it reveals is broader: a theory of accountability in which a jury verdict, a judicial sentence, and two presidential rejections are not terminal outcomes. They are the current state of a negotiation that SBF believes is still open.

FTT closed at $0.35. The market, at least, agrees with him on that last point.

As of late May 2026, Alphabet’s stock is up 23.1% year to date. Amazon’s stock is up 16.4%. Microsoft’s stock is down 12%. According to Bloomberg, Microsoft is the single biggest drag on the S&P 500’s 8.3% gain for the year — at Microsoft’s market cap weighting, its decline is more damaging to the index than the combined negative contributions of the next several worst performers.

All three companies are in the same industry. All three reported strong cloud revenue in Q1 2026. All three are spending at historically unprecedented levels on AI infrastructure. Amazon committed $200 billion in capex for 2026. Microsoft guided to $190 billion. Alphabet spent $35.67 billion in Q1 alone. The macro environment — tariff-driven inflation concerns, rate uncertainty, an equity market that briefly hit all-time highs on Iran ceasefire news — applies to all three equally.

The 35-point divergence between Alphabet’s performance and Microsoft’s is not macro. It is specific. It is structural. And the market has been articulating exactly what it is pricing, clearly enough, for anyone who has read the Q1 earnings calls and the subsequent analyst notes.

This article is the case for what that is.

The Index Weight Makes This Consequential

Microsoft’s market capitalisation places it among the two or three largest constituents of the S&P 500. When a company of that size declines 12% while the index rises 8.3%, the index is performing despite Microsoft, not with it. Bloomberg’s characterisation — that Microsoft is the market’s biggest drag — is a mathematical statement about weighted contribution, not rhetoric.

This matters for investors who hold index funds or market-weight exposures: they are long Microsoft’s AI infrastructure spending problem whether they know it or not. It matters for Microsoft itself because institutional investors who set allocations actively are asking whether the index-weight justification for holding Microsoft stock still applies when its AI thesis is underperforming its directly comparable peers by more than 35 percentage points.

It also matters for the thesis this series has been building. The financial mathematics of Microsoft’s capex vs Copilot monetisation — $190 billion against a $7.2 billion Copilot run rate, producing a 6-8 year recovery timeline — established the internal logic of why the stock should underperform. The peer comparison data confirms that the market has arrived at a version of the same conclusion, and that it is acting on it, in real time, with real capital.

The Spending Comparison That Changes the Narrative

The standard defence of Microsoft’s stock position is that the company is spending heavily on infrastructure that will generate returns over a multi-year period, and that investors who sell now are giving up the option value. This argument is correct as an abstract description of long-cycle infrastructure investment. It fails because it does not differentiate between Microsoft’s spending and Alphabet’s or Amazon’s spending, which are subject to the same abstract description and are producing sharply different market outcomes.

In Q1 2026, Amazon spent $44.2 billion on capital infrastructure. Alphabet spent $35.67 billion. Microsoft spent $30.88 billion. Amazon spent the most. Microsoft spent the least of the three. Yet Amazon’s stock is up 16% and Microsoft’s is down 12%.

The volume of spending is not the variable that is driving the differential. If it were, Amazon would be the underperformer. What the market is pricing is the expected return on that spending — and specifically, the degree to which each company’s AI infrastructure investment has a clear, near-term, monetisation pathway.

Amazon’s pathway is explicit. Its custom AI chips — the Trainium and Inferentia families — generate an annual revenue run rate already exceeding $20 billion and growing at triple-digit year-over-year rates. The valuation of Amazon’s custom silicon business alone is estimated at approximately $50 billion. AWS Bedrock has positioned itself as the neutral AI platform, offering access to every major model — Anthropic’s Claude, OpenAI’s GPT-5.4, and others — without forcing customers to commit to a single provider’s AI stack. Enterprise customers who want to hedge their AI model exposure have a natural home in AWS, and Amazon earns platform economics on whichever model wins.

Alphabet’s pathway is similarly concrete. Its TPU 8 training chip delivers three times the processing power of its prior generation. Its TPU 8i inference chip delivers 80% better performance per dollar than the generation it replaces. These are not aspirational specifications — they are the cost structure that determines what Google Cloud charges for AI workloads versus what Azure charges. Google Cloud grew at 30% in Q1 2026, taking market share from 12% to 14%, the most significant share gain among the three hyperscalers. Google Workspace AI is bundled into the productivity suite that competes directly with Microsoft 365. If Workspace AI is converting enterprise users more effectively than Copilot, the cloud-level economics reflect that within one to two quarters.

What Azure’s Numbers Actually Show

Azure grew at 39-40% in Q3 FY2026. This is not a weak number. Azure is the second-largest cloud platform globally with 21% market share, up from 20% the year prior. The infrastructure business is functioning. The supply constraint problem — Azure has been unable to meet demand because GPU provisioning is taking longer than contracted customer timelines — is being worked through, with new data centre capacity coming online throughout 2026.

The problem is not Azure. The problem is that Azure’s strength does not compensate for the product layer that sits on top of it. The Code Red designation that Nadella applied internally to Copilot’s adoption trajectory reflects this precisely. Azure is the platform. Copilot is the product. Enterprise customers who buy Azure for general cloud infrastructure are a different buyer profile from enterprise customers who are supposed to be upgrading to Copilot as their primary AI tool. The Copilot conversion story — 3.3% of the addressable Microsoft 365 base paying for it, 64% of provisioned seats going unused — is not an Azure story. It is a product-market fit story at the layer above Azure.

Bloomberg’s post-earnings summary was specific on this point: Microsoft’s April quarterly report showed “underwhelming growth in Azure cloud computing business, especially relative to Alphabet and Amazon, which suggests that peers see greater AI traction.” The phrase “greater AI traction” is analyst shorthand for the product layer. Amazon’s AI revenue, Alphabet’s Workspace AI seat expansion, and the customer migrations they are driving are “AI traction.” Azure’s growth, against a backdrop of acknowledged supply constraints, is infrastructure capacity — necessary but not sufficient to sustain the multiples that AI-era tech companies need to justify.

The Custom Silicon Gap and Why It Compounds Over Time

Microsoft’s Maia 200 chip — its proprietary AI inference processor — is live in two major data centres and delivers what Microsoft describes as a 30% improvement in tokens per dollar compared to GPU-based inference. This is real progress. It is also, against the backdrop of what Alphabet and Amazon have built, a first-generation effort in a race that its competitors entered multiple generations ago.

Alphabet has been building custom silicon for AI workloads since the original TPU in 2015. The TPU 8 generation is the culmination of more than a decade of iterative chip design. The 80% inference-per-dollar improvement is not a single generation’s gain — it is the compounding of architectural decisions made over years. Amazon’s custom chip business, now generating $20 billion in annual revenue, reflects six years of Trainium and Inferentia development that began when AWS recognised that GPU procurement at scale was a structural cost problem that needed a custom solution.

Microsoft’s Maia 200 being live in two data centres is the beginning of that journey, not a point of competitive parity. Two data centres means Microsoft is still overwhelmingly dependent on NVIDIA GPUs for the vast majority of its Azure AI inference workloads. That dependency has two cost implications: it means Microsoft’s AI infrastructure operating costs are higher per token than Alphabet’s and Amazon’s, and it means Microsoft’s long-term infrastructure margin trajectory is less certain, because NVIDIA pricing power over Microsoft is materially greater than its pricing power over two hyperscalers that have already built credible in-house alternatives.

The OpenAI Dependency: From Asset to Liability

Microsoft’s AI product strategy has been built on the OpenAI relationship. Copilot runs on GPT models. Azure OpenAI Service — one of Azure’s fastest-growing product lines — provides enterprise access to GPT-4 and its successors through the Azure infrastructure layer. The OpenAI bet was, in 2022 and 2023, among the most consequential strategic decisions in the technology industry. Microsoft moved faster than any other hyperscaler to embed a frontier model provider into its product stack.

The problem is that the relationship has evolved in ways that dilute the exclusivity thesis. OpenAI’s models are now available through AWS Bedrock. GPT-5.4 is in limited preview on AWS, with GPT-5.5 arriving within weeks. The neutral platform that Amazon has constructed — where enterprises can access Claude, GPT, and other frontier models without committing to a single cloud provider’s platform — directly competes with the proposition that Microsoft’s Azure OpenAI Service previously had near-exclusive access to build.

The non-exclusive nature of the Microsoft-OpenAI commercial arrangement has always been a known risk. The Microsoft AI squeeze dynamic — where Microsoft’s leverage over OpenAI has been eroding as OpenAI’s commercial independence has grown — anticipated this erosion. What has happened is that the erosion has accelerated faster than the model that justified Microsoft’s valuation premium assumed. When the argument for owning Microsoft over Alphabet or Amazon was partly “they have the most direct pipeline to the best AI models,” and then those models become available on AWS, part of the valuation differential evaporates.

Microsoft still has meaningful advantages from the OpenAI relationship: priority access to model updates, infrastructure integrations that run through Azure, and the Microsoft 365 Copilot embedding that places GPT models inside the productivity applications that enterprise workers use daily. These are real. But they are no longer exclusive. And in a market where Alphabet has built its own competitive models (Gemini) and Amazon offers a multi-model neutral platform, “no longer exclusive” matters more than it did two years ago.

The Valuation Paradox

Microsoft trades at 24.4 times forward earnings. Amazon trades at 34.2 times. Alphabet trades at 34.9 times. The company with the weakest custom silicon position, the most product-layer adoption problems, and the diluting partnership exclusivity trades at a 30% discount to its direct peers.

Some of this discount is structural and appropriate. Microsoft’s revenue base is more mature than Amazon’s, which is still in a high-growth phase across AWS and e-commerce. Microsoft’s earnings are higher-quality in the short term — it generates substantial free cash flow — which compresses the multiple that growth-dependent investors assign. These are legitimate valuation considerations that have always applied.

What is new in 2026 is that Microsoft’s forward earnings multiple has compressed relative to where it traded in 2024 and 2025. The compression encodes the market’s reassessment of Microsoft’s AI growth trajectory. When the consensus was that Microsoft’s OpenAI relationship, Copilot bundle, and Azure scale would produce AI-driven earnings acceleration, the stock commanded a premium. As the Copilot adoption data accumulated — 3.3% penetration, 64% seat utilisation, ChatGPT preferred by enterprise users at 76% vs Copilot at 18% — and as the Azure growth showed supply constraints rather than demand-driven acceleration, the premium has become a discount.

Stifel’s February 5 downgrade — rare for an analyst covering a company with Microsoft’s market standing — made this arithmetic explicit. Brad Reback cut his price target from $540 to $392 and moved his rating to Hold. His FY27 capex estimate of $200 billion, against a Street consensus of $160 billion, implied that the spending acceleration would compress margins further before any monetisation uplift materialised. His gross margin forecast for FY27 of 63% against a consensus of 67% is not a small difference — it is four points of margin on a company generating hundreds of billions in revenue. The Stifel note did not create the discount. It formalised it in institutional language that other analysts have subsequently echoed.

The Microsoft 365 Defence and Its Limits

The bull case for Microsoft that is still being made by its defenders runs through Microsoft 365 rather than Copilot specifically. The bundling strategy that Microsoft has deployed — progressively embedding Copilot features into standard Microsoft 365 tiers at price points that make standalone Copilot pricing feel unnecessary for many customers — is a real strategic response to the adoption problem. If Copilot cannot convert as a premium add-on, make it a baseline feature and recover the economics through bundle price increases.

The limit of this defence is that it works only if Microsoft 365 itself retains its enterprise foothold as Google Workspace AI traction grows. If Google Workspace’s AI capabilities improve to the point where the switching costs from Microsoft 365 to Google Workspace become acceptable for a meaningful segment of enterprise customers, the bundle strategy loses its moat. Google Cloud’s 14% market share, up from 12%, is a data centre and workload statistic — but it is also directionally consistent with enterprise IT departments that are re-evaluating their Google vs Microsoft footprints and finding the Microsoft story less compelling than it was three years ago.

The split between AI capex spenders and the rest of the S&P 500 was always going to require differentiation within the capex-spending cohort. Not every company that spends on AI infrastructure will generate comparable returns. Microsoft’s position in that differentiation — as the largest spender with the weakest product-layer monetisation story — is the reason the market has applied a discount that its peers have not received.

The Counterargument: Why Some Analysts Are Still Buyers

The case for Microsoft as a value-at-current-price argument has reasonable foundations. At 24.4 times forward earnings, a company generating the free cash flow that Microsoft generates, with the enterprise installed base it maintains, is not obviously expensive on a long-term hold basis. Barchart noted that Microsoft stock is up nearly 30% from its March 2026 lows — a recovery that suggests institutional buyers at lower prices exist and have been active.

The structural arguments: Azure supply constraints are temporary, and when the capacity backlog clears, growth should accelerate. Copilot adoption is a long cycle — enterprise software has historically taken 18-36 months to reach meaningful penetration after initial rollout — and 3.3% penetration at two years after launch is not necessarily a ceiling. Microsoft’s Personal Computing segment, down 1%, may bottom as the PC replacement cycle turns. And the Maia 200 chip in two data centres is the start of a multi-year custom silicon programme that could eventually produce the same infrastructure cost advantages that Alphabet and Amazon enjoy today.

These arguments are not wrong. They are arguments about a future in which Microsoft’s current problems are transitional rather than structural. The difficulty is that the same argument — “this is transitional, wait for the product cycle to turn” — has been the Microsoft bull case for the better part of two years, while Copilot penetration has not materially accelerated and the peer performance gap has continued to widen.

At some point, the distinction between “transitional problem” and “structural problem” is decided by evidence, and the evidence that would confirm the transitional read — accelerating Copilot conversion, improving seat utilisation, positive feedback from enterprise deployments, Maia-enabled margin improvement — has not yet arrived in the numbers. Until it does, the discount the market is applying reflects an appropriate Bayesian update, not an overreaction.

What Would Change the Thesis

The conditions under which Microsoft’s valuation discount narrows relative to Alphabet and Amazon are specific and identifiable. They are not speculative — they are testable claims about outcomes that will either appear or not appear in the next two to four quarterly earnings reports.

First: Copilot penetration acceleration. A move from 3.3% to 8-10% of the addressable Microsoft 365 base on paid Copilot plans, within four quarters, would represent a product-market fit inflection. The seat utilisation metric — currently 36% of provisioned Copilot seats in active use — would need to climb above 60% to signal that the adoption problem is being resolved rather than managed. These numbers are not visible in the current data.

Second: Maia 200 at scale. Microsoft’s custom chip is in two data centres. At ten or more, with disclosed economics that demonstrate inference cost parity with Alphabet’s TPU 8i performance per dollar, the custom silicon dependency on NVIDIA becomes a story about maturation rather than structural disadvantage. A specific management disclosure on the Maia 200 deployment roadmap, with dates and capacity commitments, would move this from aspiration to plan.

Third: The OpenAI relationship crystallising. A refreshed commercial agreement that establishes the terms of the Microsoft-OpenAI partnership through the mid-2030s — with explicit protections against further third-party distribution that dilutes Azure’s model-access advantage — would resolve the platform risk that the AWS Bedrock GPT availability introduced. Without that crystallisation, the partnership’s value continues to erode.

None of these are scheduled announcements. Q4 FY2026 earnings, expected in late July, will provide the next substantive data points on Azure growth and Copilot adoption. If the Copilot penetration number in that report does not show meaningful improvement from the 3.3% figure that has defined the story since early 2026, the market’s discount will not narrow — it will widen.

The Synthesis

Microsoft is not failing. Its infrastructure business is strong. Its enterprise relationships are durable. Its free cash flow generation is exceptional. The company will not collapse, and the people predicting its irrelevance have consistently overestimated how fast enterprise technology transitions happen.

What Microsoft is doing is underperforming the specific version of itself that the market priced in 2024 — the AI-accelerated growth story in which Copilot converts enterprise users at scale, the OpenAI relationship provides durable product differentiation, and Azure’s infrastructure spending produces returns that justify a premium multiple against Alphabet and Amazon.

That version of Microsoft has not arrived. In its place is a company with a supply-constrained cloud business, a product-layer adoption problem that has persisted across multiple remediation attempts, a custom silicon programme that is two or three generations behind its best-in-class peers, and a flagship AI product that enterprise users prefer less than its primary competitor in a direct preference survey at a ratio of 76% to 18%.

The broader enterprise AI spending accountability reckoning was always going to differentiate between companies whose AI investments converted and companies whose did not. Microsoft, at the moment, is the most expensive exhibit in that reckoning — not because it has failed in any terminal sense, but because it is the company that has spent the most institutional credibility on an AI transition story that the product numbers have not yet confirmed.

Alphabet is up 23%. Amazon is up 16%. Microsoft is down 12%. The market is not confused. It is doing its job.

What the Stock Divergence Reveals About Who Has the AI Product Right

Julie Zhuo’s framework for understanding product divergence begins with the jobs users hire each product to do. When three companies in the same industry, with roughly similar AI investment levels, produce wildly different stock returns in the same period, the divergence is not random. It is the market’s assessment of which company’s product is actually accomplishing the job the user needs done, versus which company is spending on AI capability that has not yet connected to a job users are hiring it for. The Alphabet/Amazon versus Microsoft performance gap is a product gap masquerading as a financial gap.

Alphabet’s position is the most legible. Google Search has a job that billions of people hire it to do multiple times per day. AI-enhanced search — whether through AI Overviews, Gemini integration, or deeper result synthesis — improves the performance on that existing job without requiring users to adopt a new workflow. The adoption friction is near-zero because the surface is the same and the job is the same. Alphabet is adding AI capability to a product users already trust to accomplish a job they already have. The market is pricing that compounding correctly.

Amazon’s position is equally legible from a different angle. AWS is infrastructure. When AWS adds AI capability — SageMaker improvements, Bedrock model access, Trainium chips — the customers adopting those capabilities are developers and enterprise IT teams who are specifically hired to evaluate and deploy new technical capabilities. The job-to-be-done is “make our AI infrastructure more capable.” The customer is predisposed to adopt improvements to the infrastructure they are already managing. Amazon is adding AI capability to a market that already knows it needs more AI capability. Again, near-zero adoption friction for a well-defined job.

Microsoft’s Copilot problem is a job-to-be-done problem. Enterprise AI adoption at 3.3% Copilot penetration means that 96.7% of the potential enterprise user base has the capability available and is not using it regularly enough to generate measurable productivity gain. The job Copilot is positioned to do — “make every knowledge worker more productive across all their workflows simultaneously” — is too broad to be a single coherent job. Users who are highly productive at a specific workflow do not have a generic productivity problem. They have specific friction points in specific workflows. Copilot’s value is highest when it addresses a specific friction point that the user encounters frequently enough to justify the habit change required to use AI assistance for it. Friction is the silent churn driver, and the Copilot adoption data suggests that the friction of adopting AI assistance is not yet lower than the friction of the existing workflow it is designed to replace.

The Microsoft developer platform dynamic illustrates the second dimension of the divergence: Microsoft has been extracting margin from its developer ecosystem at the same time that it has been asking that ecosystem to adopt new AI-assisted workflows. Those are competing signals. A developer who is paying more for GitHub, more for Azure, and more for Microsoft 365 while also being asked to allocate time to learning Copilot integration has a rational skepticism about whether the new capability is genuinely for their benefit or is primarily another margin-extraction vehicle. That skepticism slows adoption in a way that pure capability arguments cannot overcome, because the problem is not capability — it is trust.

The product correction that the stock divergence is signaling is specific: Microsoft needs to identify the three to five workflows where Copilot dramatically reduces friction for users who encounter those workflows daily, focus adoption efforts there, demonstrate measurable outcomes in those workflows, and use those demonstrated outcomes to expand adoption to adjacent workflows. That is the Zhuo approach: narrow the job, demonstrate the result, expand from the result. The corporate capital return environment makes this more urgent — companies that are buying back stock rather than investing in productivity tools are signaling that the expected return on productivity investment is below their cost of capital. If Copilot cannot demonstrate the narrow outcome that justifies the broad investment, the market will keep pricing the gap. Prediction markets on Copilot paid seat growth through year-end are pricing a positive trajectory but a slow one — which is exactly what the job-to-be-done analysis predicts for a product that hasn’t yet found its narrowest, most defensible use case.

Microsoft’s third-quarter FY2026 earnings, released April 30, were unambiguously strong by the metrics that have historically moved the stock. Revenue of $82.89 billion came in above consensus. Revenue growth of 18 percent year over year was the fastest in several quarters. Azure cloud revenue grew 40 percent year over year — an acceleration from prior quarters. Intelligent Cloud, the segment that includes Azure, contributed $26.75 billion. Operating income was up 16 percent. Every major business unit beat estimates. The call was, by traditional earnings analysis, a strong quarter.

The stock fell 5 percent on the day. Microsoft entered 2026 at around $430 per share and is down 15.7 percent year to date against an S&P 500 that has climbed to record highs. The stock’s underperformance relative to the index in a record-setting year is not explained by earnings misses, margin compression, or revenue deceleration. It is explained by something the earnings release documents clearly and the sell-side comps less readily: the gap between what Microsoft is spending on AI infrastructure and what the product that is supposed to generate a return on that infrastructure is currently producing.

That product is Copilot. That gap is the subject of this analysis.

The Monetization Math

Microsoft’s 2026 capital expenditure guidance is $190 billion — a 61 percent increase from 2025, and more than three times the company’s capex in 2024. The company’s own communications have been clear about what this spending is for: AI infrastructure. Data centre construction, GPU procurement, and the networking and power infrastructure required to run large-scale AI inference and training workloads. The $190 billion commitment is not speculative. It is a multi-year programme with supplier contracts, construction permits, and public disclosure in Microsoft’s forward guidance.

For that commitment to generate an acceptable return, Microsoft needs revenue growth that exceeds the capex increase over a reasonable horizon. The primary mechanism for that revenue growth, in Microsoft’s stated strategy, is Copilot: the AI layer integrated into Microsoft 365 and the broader Office product suite, priced at $30 per user per month as an add-on to existing M365 subscriptions. The thesis is straightforward: enterprise customers are already paying for M365 at scale; Copilot converts that installed base into a higher-revenue-per-seat business while the AI infrastructure investment enables the product’s capabilities.

The adoption data measures how that thesis is performing against the base case required. Independent research published in early 2026 found that approximately 3.3 percent of Microsoft’s commercial M365 subscriber base has converted to paid Copilot. Microsoft has disclosed 20 million paid Copilot seats as of April 2026, up from 15 million the prior quarter. Against Microsoft’s addressable commercial M365 base of more than 450 million users, 20 million represents 4.4 percent at the high end of the range — consistent with the 3.3 percent figure from independent research, given definitional differences in how “addressable base” is counted.

At 20 million paid seats and $30 per user per month, Copilot’s current annual revenue run rate is approximately $7.2 billion. That is a meaningful number in absolute terms and represents genuine growth — 15 million seats three months earlier implied a $5.4 billion run rate, so the business is adding roughly $7 million in annual run rate per day. But against a $190 billion annual capex commitment, the ratio of current Copilot monetisation to infrastructure investment is approximately 1 to 26. For every dollar Copilot currently generates in annual revenue, Microsoft is spending $26 on the infrastructure required to support its long-term ambitions.

The capex recovery timeline depends on the adoption growth rate. If Microsoft doubles Copilot seats annually — from 20 million to 40 million to 80 million, approaching meaningful penetration of the 450 million seat addressable market by 2028 or 2029 — the monetisation trajectory begins to justify the infrastructure commitment. If the growth rate slows, or if the 3.3 percent conversion figure reflects a structural ceiling in enterprise AI adoption rather than an early-stage penetration curve, the timeline extends materially. The analyst estimates of 6 to 8 years to recoup the $190 billion capex commitment at current adoption rates are not pessimistic projections. They are arithmetic.

What the NPS and Preference Data Say

The adoption percentage is the headline figure. The preference data is the more diagnostic signal. A product with 3.3 percent penetration in a rapidly expanding market might still be on the right adoption trajectory if users who have it love it and word-of-mouth is building toward the broader base. That is how early enterprise software products grow: slow initial uptake, high satisfaction among early adopters, organic expansion through intra-organisation advocacy.

Copilot’s satisfaction data does not support that reading. The product’s accuracy Net Promoter Score — the measure of whether users would recommend it — stood at negative 19.8 in January 2026. A negative NPS means more users are actively discouraging adoption than promoting it. For reference, enterprise software products considered strong performers typically have NPS scores above 30. Negative NPS at the product level is not a growth story. It is a retention and advocacy problem that directly limits the organic expansion mechanism that enterprise software companies depend on for penetration growth.

The competitive preference data compounds the reading. In surveys of enterprise users who have access to both Copilot and ChatGPT, 76 percent identify ChatGPT as their primary AI productivity tool versus 18 percent for Copilot. When all AI tools are available, Copilot’s share falls to 8 percent. These are not marginal preference differences — they are dominant preference differentials in a head-to-head context that should, by the logic of Microsoft’s distribution advantage, favour Copilot. Microsoft’s product is embedded in every M365 subscription, accessible from the toolbar of every Word document, integrated into every Teams conversation. The competitor it is losing to requires a separate login and an additional subscription. The product with the better distribution is losing by a four-to-one margin.

The combination of negative NPS and 76-to-18 competitive preference is the most direct available evidence that the Copilot adoption problem is not primarily a marketing problem or an awareness problem. Enterprises that have provisioned Copilot access are choosing not to use it, or are choosing a competitor’s product when both are available. The 64 percent non-usage rate — the share of provisioned Copilot seats that see no active use — reflects the same dynamic at the usage level: the product is present in the environment and is not being adopted by the majority of the employees it is designed to serve.

What the Stock Is Pricing

Microsoft’s stock performance in 2026 is the market’s synthesis of the data above, translated into price terms. A company with 18 percent revenue growth, 40 percent cloud growth, and operating income expansion that misses no significant estimate is not typically a stock that underperforms a rising market by 20 percentage points. The market’s explanation for that underperformance is embedded in the forward multiple compression that the price action represents.

Microsoft traded at approximately 35 to 38 times forward earnings entering 2026 — a premium multiple reflecting expectations of sustained AI-driven revenue acceleration. A premium multiple compresses when the anticipated acceleration fails to materialise at the rate or on the timeline the multiple implied. The compression does not require a bad quarter. It requires only that the expected path is not being validated at the pace that justified the entry multiple. Microsoft’s Q3 beat did not validate the path at the rate required; it confirmed growth but did not demonstrate that the Copilot monetisation trajectory was closing the gap between capex commitment and revenue return fast enough to justify the prior multiple.

The five percent post-earnings decline is the most precisely available evidence of this dynamic. Investors who reviewed the earnings, the guidance, and the Copilot seat data sold the stock on a beat. That behaviour is not irrational. It reflects an updated forecast: given the current adoption metrics, the forward path to Copilot penetration that would justify a premium AI multiple is longer than the pre-earnings multiple implied. The sell is not a vote that Microsoft has failed. It is a repricing of the timeline to success.

The year-to-date underperformance extends this reading across a longer window. While the S&P 500 has set records and AI infrastructure suppliers — Nvidia, TSMC, Micron — have been repriced upward for their structural scarcity value, Microsoft has been repriced downward for its structural monetisation gap. The infrastructure suppliers are selling something scarce that is in high demand. Microsoft is selling something abundant (M365 seats) whose conversion rate to premium AI revenue is lower than the infrastructure investment requires.

The Bundling Strategy and Its Limits

Microsoft’s response to the monetisation gap has been to shift from a standalone Copilot add-on model toward a bundled inclusion model. Beginning in late 2025, Microsoft began incorporating Copilot capabilities into higher-tier M365 SKUs rather than requiring a separate $30/month purchase decision for every seat. The Copilot bundling decision reflects a specific strategic calculation: lower the per-seat barrier to access to drive adoption, even at the cost of lower per-seat revenue, on the thesis that broad adoption will validate the product value and enable subsequent price increases or higher-tier migration.

The bundling strategy is a defensible response to slow adoption. It is also a concession. A product that requires bundling to drive usage is a product that has not independently demonstrated sufficient value to command the standalone purchase decision at the price required. Enterprise software products that bundle their way to adoption can convert that adoption into pricing power over time — if the product genuinely embeds into workflows and creates switching costs. They cannot create pricing power from adoption alone if the adoption is driven by availability rather than demonstrated value.

The NPS data suggests that current Copilot users are not experiencing the product as workflow-embedding. If they were, the NPS would be positive and the ChatGPT preference differential would be narrowing. The bundling strategy addresses the access barrier. It does not address the satisfaction problem. A user who was provisioned Copilot as part of an M365 bundle and chose not to use it is, after the bundle, provisioned and choosing not to use it. The access barrier was not the binding constraint for that user. The product value was.

The Azure Counter-Argument

The most coherent bull case for Microsoft at current prices does not depend on Copilot’s consumer-level productivity suite adoption. It depends on Azure. Azure grew 40 percent year over year in Q3 FY2026, and Azure’s AI-specific revenue — the inference, training, and model-serving workloads that run on Microsoft’s infrastructure — is growing at rates above the overall Azure number. The $37 billion annual AI revenue run rate that Microsoft disclosed represents a genuinely large and rapidly growing business that does not depend on enterprise users clicking a Copilot button in Word.

The Azure bull case argues that Microsoft has already won the enterprise AI infrastructure race — that the combination of Azure’s scale, the OpenAI partnership, and the enterprise trust relationships that Microsoft has built over decades of Windows and Office deployment constitute a durable competitive position that is monetising well even if the Copilot consumer layer is slow to develop. On this reading, the Copilot adoption metrics are noise: a product-level challenge that will eventually resolve through iteration, and that does not represent a structural threat to the underlying Azure-based business that is already delivering strong results.

The broader enterprise AI spending accountability problem that is emerging across corporate America adds a dimension to this counter-argument. If enterprises are scrutinising AI ROI more carefully, the companies whose AI products can demonstrate measurable financial returns will attract more spending, and the companies whose products have a negative NPS and a preference disadvantage against free alternatives will face a harder renewal environment. Azure’s infrastructure business benefits from AI capex growth broadly — Microsoft does not need to win on the Copilot product layer to capture infrastructure spending from enterprises building AI applications on Azure. The infrastructure revenue and the product-layer adoption challenge are partially separable.

Why the Azure Growth Does Not Close the Gap

The Azure counter-argument is correct on its own terms. The question it does not fully answer is whether Azure’s growth, combined with the rest of Microsoft’s business, produces a return on $190 billion in annual capex at a rate that justifies the infrastructure investment. Azure’s 40 percent growth from a base of approximately $65 billion annually represents incremental revenue of roughly $26 billion in the current year. If Azure continues compounding at 40 percent annually — a rate that has been sustained for several quarters but that will face comparison difficulty as the base grows — the cumulative Azure revenue over five years is substantial.

The challenge is that the $190 billion capex figure is not solely for Azure. It funds the infrastructure that supports Copilot’s consumer-layer use cases, Microsoft’s consumer AI products, and the broader capacity expansion that the company has committed to. The marginal return on the incremental infrastructure investment depends on what that infrastructure enables — if it enables Azure workloads at the current growth rate, the return case is defensible. If it enables Copilot workloads at the current adoption rate, the return case requires a longer horizon. The capex is fungible; the revenue return is not.

Hamilton Helmer’s distinction between process power and scale economies is useful here. Azure’s growth reflects genuine scale economies — a larger infrastructure base serves more customers at lower marginal cost, and enterprise trust in Azure’s reliability compounds through multi-year contracts and integration depth. That is a durable competitive position. Copilot’s challenge is a process power problem: the product has not yet embedded deeply enough into enterprise workflows to create the switching costs that would justify the premium pricing and predict durable adoption. Process power accrues from actual workflow embedding, not from bundled availability. Until Copilot generates positive NPS and closes the competitive preference gap with ChatGPT, the process power thesis is unproven.

The operational response Microsoft has taken — Nadella’s Code Red designation, the leadership restructure, the AI team ring-fencing in the voluntary buyout — is proportionate to the urgency of the problem. A chief executive taking personal ownership of a product’s adoption curve, restructuring the leadership team, and restructuring the workforce to free up capital for infrastructure investment is not a company unaware of its situation. These are the right diagnostic and operational responses to the challenge the adoption data describes.

The Structural Question the Earnings Metrics Cannot Answer

The earnings call format is designed to report on the quarter. It is not designed to address the structural question that the adoption data poses: what is the equilibrium penetration rate for Microsoft Copilot in a market where a well-funded, highly capable competitor is available at comparable cost and is preferred 4-to-1 by users who have tried both?

That question matters more than the Q3 beat because it determines the long-term revenue trajectory that justifies the $190 billion capex commitment. If the equilibrium penetration rate is 25 to 30 percent of the M365 addressable base — the rate that would produce Copilot revenue sufficient to justify the infrastructure investment over a reasonable horizon — then the current 4.4 percent represents an early-stage position with substantial runway, and the NPS and preference data represent solvable product challenges. If the equilibrium rate is closer to the current range — a segment of the market that finds genuine value in AI assistance within the Office environment, but a smaller segment than the addressable market total — then the capex commitment is sized for a scenario that the product data does not support.

The honest answer is that the equilibrium penetration rate is not knowable from current data. The market is attempting to price it through the stock’s forward multiple compression. The NPS data and the competitive preference data are the best available proxies for where the equilibrium is heading. Neither of them is pointing toward the 25-to-30 percent scenario that the capex commitment requires. They are pointing toward a product that needs to improve materially before the adoption curve bends upward at the rate required.

What Would Change the Math

Three developments could change the monetisation calculus meaningfully within the planning horizon the market is pricing. The first is a product breakthrough in Copilot’s utility that converts the NPS from negative to materially positive and narrows the ChatGPT preference gap. This is the product iteration path — the path that Nadella’s Code Red response is pursuing. The signal to watch is not the seat count, which can be grown through bundling without reflecting genuine utility. It is the NPS trajectory and the competitive preference data over the next two to three quarters.

The second is the agentic AI transition that Jensen Huang has argued will drive a tenfold increase in compute demand above the generative AI baseline. If agentic AI — autonomous multi-step task completion that replaces human workflow execution rather than merely assisting it — becomes the dominant enterprise use case for AI in 2027 and 2028, Microsoft’s infrastructure investment is correctly positioned for the demand it will serve. Agentic AI deployed through enterprise workflows would produce measurable productivity economics that the current Copilot productivity assistance model cannot replicate: cost per task completed rather than monthly subscription per seat. The return model shifts from adoption rate to task volume, and task volume scales differently than seat count.

The third is the competitive narrowing that Microsoft’s distribution advantage should theoretically produce over time. OpenAI’s ChatGPT is a consumer and SMB product that competes with enterprise Copilot at the feature level but lacks Microsoft’s depth of integration into the enterprise Office environment. If Microsoft solves the product quality gap — the NPS problem — the distribution advantage in enterprise workflows could convert into penetration at rates that the current preference data does not reflect. The distribution advantage is real. It is not currently being expressed in preference outcomes. The question is whether the product iteration resolves that gap before the competitive field shifts further.

The Honest Assessment

Microsoft in 2026 is a company that has made the right strategic bet at the right time — AI infrastructure investment at scale, before the demand curve fully materialised — and is now in the period where the investment is visible and the return is not yet proven. That period is uncomfortable for investors who price on forward multiples. It is not, by itself, evidence that the strategy is wrong.

The discomfort is specific and quantifiable. A $190 billion annual capex commitment requires a monetisation path at scale. The current Copilot adoption data — 4.4 percent penetration, negative NPS, 4-to-1 preference disadvantage against the primary competitor — does not yet describe that path. The Azure growth provides a strong supporting case, but the Azure revenue base is not the capex beneficiary on the scale the infrastructure programme requires. The market is pricing the gap between the commitment and the demonstrated path to return. The gap is real, and the metrics that would close it have not yet moved in the required direction.

The prior analysis on this site — the Code Red designation, the Xbox and Activision reckoning, the voluntary buyout’s capital allocation purpose — described a board that has made the correct diagnosis and is taking minimally sufficient action. The monetisation math adds a sharper edge to that reading. Minimally sufficient action in an operational context is a programme. Minimally sufficient action against a $190 billion annual capex commitment with a six-to-eight-year recovery horizon at current adoption rates is a countdown. The programme needs to work. The timer is running.

The 7 Powers analysis of Copilot’s strategic position is more complicated than the adoption figure implies, and the complication cuts in both directions. Counter-positioning is the clearest current advantage: Google Workspace Gemini and Salesforce Einstein are the only alternatives with comparable corpus access to enterprise data, and neither carries the M365 integration depth that makes Copilot’s context window substantively useful rather than performatively intelligent. The switching cost power is real but conditional — it compounds only when Copilot is genuinely embedded in workflow rather than sitting adjacent to it, which is exactly the distinction the 3.3 percent figure does not resolve. That percentage could represent deep integration among early adopters who have reorganized document and communication workflows around the tool, producing durable switching costs and a defensible installed base. Or it could represent shallow usage among a cohort that has not meaningfully changed how work gets done, producing near-zero switching costs and a renewal risk that does not appear in the current contract metrics. Scale economies are unambiguous: AI inference at Azure’s capacity produces a lower marginal cost per session than any independent competitor can approach. But scale economies compound only if the user base that benefits from them keeps growing. The question the earnings metrics cannot answer is whether the powers already built are sufficient to fund the growth that the other powers require. The timer does not wait for the answer.