On 10 June 2026 the European Commission put out a near-final draft of the rulebook that tells anyone shipping AI-generated images, audio, video or text inside the EU how to mark it so a machine can tell it apart from human work. It is called the Code of Practice on marking and labelling of AI-generated content, and it spells out the practical detail behind Article 50 of the AI Act, the transparency clause that becomes binding on 2 August 2026. In plain terms: if your bank's chatbot writes an email, your carmaker's infotainment voice answers a driver, or your agency's tool generates a campaign visual, that output now has to carry a hidden tag that downstream platforms can read. The Code arrives a week after the EU's broader Technological Sovereignty Package, which bundles in a Cloud and AI Development Act. Marking is no longer optional polish.

In a glass-walled press room above Rue de la Loi, Lucilla Sioli, director of the EU AI Office, walked reporters through a slide that did something unusual for a Brussels deck: it showed code. Not legal code, but a snippet of C2PA-style manifest data, the kind of cryptographic metadata that travels with a file from the moment a model spits it out. “The principle is simple,” she said, paraphrasing the room's mood more than reading from her notes. “If a machine made it, a machine has to be able to say so.” Behind her, a slide listed the four mandatory layers the Code now formalises: digitally signed provenance metadata, imperceptible watermarks robust to crops and re-encodes, generator-side fingerprinting and logging, and detection protocols that downstream platforms can call. That was the choreography of 10 June 2026. The substance had been building for half a year. The first draft, published just before Christmas 2025, ran shorter and softer; the March revision absorbed more than 500 stakeholder comments; this June text reads like an engineering brief. Signatories — the major model providers who voluntarily commit — will be listed publicly in July, mirroring the pattern set last summer when 24 firms put their names to the General-Purpose AI Code on a similar timeline. The Code does not stand alone. One week earlier, on 3 June, Executive Vice-President for Tech Sovereignty Henna Virkkunen rolled out the European Technological Sovereignty Package, a bundle that includes a Chips Act 2.0, an Open Source Strategy and the new Cloud and AI Development Act (CADA). “We live in a world where geopolitics and technology are inseparable,” Virkkunen told the CNBC camera scrum that followed. “It is time for Europe to be in control of its data, of its supply chains, and of its future.” CADA defines four sovereignty assurance levels for cloud and AI services, aims to triple EU data-centre capacity within five to seven years, and gives public bodies a single rulebook to procure against. Read together, the two files describe a Europe that wants both the substrate (sovereign infrastructure) and the surface (provenance-tagged output) under its own jurisdictional control. For enterprise teams the practical date is 2 August 2026. That is when Article 50 obligations kick in for providers of generative AI systems and for the deployers who push their output into the world. Compare it to the GDPR ramp: that regulation passed in April 2016 and applied in May 2018, giving firms 25 months of runway. Article 50 has effectively given marketing, comms and product teams 51 days from this Code's publication. “The runway,” as one Frankfurt CISO put it on a private channel, “is a taxiway.” Not by accident: the Commission wants the rulebook locked before the obligation lands, so signatories can claim a presumption of conformity rather than argue case by case.

Strip away the legal preamble and the Code prescribes four interlocking layers. The first is digitally signed metadata, in practice the C2PA manifest format that Adobe, Microsoft, Google, Leica, Nikon and OpenAI have spent two years stitching into cameras, editors and model APIs. A signed manifest records who generated the asset, with what tool, when, and whether human edits were declared. It is tamper-evident: break the seal and downstream verifiers can tell. The second is imperceptible watermarking inside the pixels or audio waveform itself — Google DeepMind's SynthID is the reference implementation cited in stakeholder workshops — designed so that even a cropped fragment carries enough signal for detection. The third is fingerprinting and content logging on the generator side, an optional layer that lets a provider check whether a suspect file came from its model. The fourth is a detection and verification protocol that platforms, broadcasters and fact-checkers can query. This multi-layered choice is deliberate and contested. Article 50(2) of the AI Act demands marking be “effective, interoperable, robust and reliable” — four adjectives that, as an arXiv paper from researchers at TU Munich noted in March 2025, no single technique currently satisfies. Watermarks survive re-encoding but can be degraded; metadata is precise but strips out the moment a screenshot is taken; fingerprinting works well at scale but only if the original generator cooperates. The Code's answer is to stack them. The Computer & Communications Industry Association, the lobby that represents Google, Meta, Amazon and most of the big foundation-model players, has called the requirements “burdensome” and warned of “banner blindness” if every AI-touched pixel triggers a user-facing notice. The Information Technology Industry Council went further in a TechWonk post, arguing that prescriptive technical mandates risk freezing a still-evolving research field. The catch is what counts as a “deepfake”. The AI Act defines it as image, audio or video that resembles real persons, objects, places or events and would likely mislead a viewer into believing it is authentic. That sweeps in everything from a synthetic ad voiceover to a customer-service avatar to a re-aged car commercial. Deployers — not just model providers — must “clearly and distinguishably” disclose. For a DAX40 marketing team that means two layers of obligation: the agency using Midjourney or Runway has to ship machine-readable provenance, and the brand publishing the campaign has to add a human-readable label. For text on “matters of public interest,” publishers face a parallel disclosure rule, with carve-outs for content that has been substantively edited by a human under editorial responsibility. The historical analogy is GDPR, but the closer parallel is the U.S. Volcker Rule. Volcker, at adoption, ran to roughly 950 pages of preamble and rule text per major bank — a document so dense that compliance teams spent years mapping their trading books against it. The AI Act plus the marking Code is shaping up similarly: not by page count, but by the depth of plumbing changes it forces. Every generative pipeline that produces customer-facing output across an EU operation has to gain a provenance hook, a watermark insertion point, and a logged audit trail. Most do not have one today. And because CADA arrived in the same fortnight, the question of where those pipelines run also got harder. CADA's four assurance levels (from baseline self-attestation up to fully EU-controlled cloud) will shape procurement at every Bundesministerium and, through public-sector demand signals, at the regulated industries that sell into them.

On 10 June 2026 OpenAI and Oracle published a short joint note: from the coming weeks, any Oracle customer can spend existing Oracle Cloud Universal Credits on OpenAI frontier models and the Codex coding agent through Oracle Cloud Infrastructure. No new vendor onboarding. No new SaaS contract. No fresh security review. For a Fusion ERP buyer in Wolfsburg or an Allianz finance team in Munich, the OpenAI API becomes a line item against a cloud commitment they already signed. It is the second time in the history of generative AI that a hyperscaler has welded OpenAI directly onto its enterprise distribution rails. The first time was Microsoft. This one matters because it lands on a different installed base: ERP, healthcare, and database. The procurement layer just became the AI distribution layer.

The press release went up on openai.com at the end of the New York trading day, an unsigned company note dated 10 June 2026 and tagged Partnerships, 2026, Codex, API Platform. No event, no joint keynote, no shareholder call. Two paragraphs and a line of fine print: “Contact your Oracle sales representative for details, timing, and availability.” Inside Oracle Park in Redwood Shores, the news travelled the way enterprise software news usually does — through a Slack channel of field reps who had been pre-briefed days earlier and whose pipeline reviews had already been rewritten around the offering. The mechanism is mundane and that is the point. Oracle Cloud Universal Credits — known internally as UCM — are the prepaid currency of OCI: a pool of money committed up front, drawn down as customers consume Autonomous Database, compute, storage, MySQL HeatWave, or any other OCI service. From the coming weeks, eligible credits can also be drawn down against calls to OpenAI frontier models and Codex. The OpenAI post described the goal as letting enterprises “deploy AI through the procurement processes and governance frameworks they already trust” and “move from AI ambition to production impact.” There was no joint pricing schedule, no rev-share number, and no mention of which exact model tier — GPT-5.5, GPT-5.4, the open-weights gpt-oss family, or all of them — would be available at launch. The corporate context is what gives the footnote its weight. In July 2025 OpenAI signed what is now described as a $300 billion, five-year cloud agreement with Oracle, starting in 2027, to develop up to 4.5 gigawatts of additional Stargate capacity — capacity Oracle CEO Clay Magouyrk casually defended on CNBC last October with the phrase, “of course OpenAI can pay $60 billion per year.” In April 2026, Microsoft and OpenAI publicly renegotiated their relationship, ending the Azure exclusivity that had defined enterprise GenAI distribution since 2023; Microsoft kept a non-exclusive licence through 2032 and a capped revenue share, but lost the right to be the only hyperscaler reselling OpenAI's models. Oracle was the first to convert that change into a procurement product. AWS Bedrock followed within weeks. Google has so far stayed on Gemini. The pivot is this. The Microsoft–OpenAI relationship was built on capital and compute: Azure was where the models lived because Microsoft had paid for them to live there. The Oracle relationship has the same compute spine — Stargate, the 4.5 GW build, the contracted orders that pushed Oracle's remaining performance obligations past $523 billion this spring. But the 10 June announcement is not about compute. It is about the demand side: the millions of seats inside Oracle Fusion ERP, NetSuite, Cerner, JD Edwards and E-Business Suite that already have a buying relationship, a security review, a data-processing addendum, and a renewal date. OpenAI just rented that distribution channel. As one Oracle field engineer put it on LinkedIn within hours of the post: “Customers don't have to argue with procurement anymore. The argument is already won.”

To understand why a clause about prepaid credits is the most consequential AI announcement of June, look at what failed to scale in 2024 and 2025. The bottleneck in enterprise GenAI adoption stopped being model quality eighteen months ago. GPT-4 was good enough for almost every realistic enterprise workload before the Microsoft–OpenAI exclusivity ended; GPT-5.5 is, by any honest read, well past it. The bottleneck became procurement: vendor risk reviews that take six to nine months at a German bank, data-processing addenda that have to be renegotiated when the data leaves an existing cloud region, an information-security questionnaire that demands answers about every sub-processor down the chain. Every one of those steps is a six-figure consulting line and a calendar quarter of delay. The historical comparison is Salesforce on AWS. From 2016 onwards, Salesforce ran the bulk of its public cloud workloads on AWS — but more importantly, AWS Marketplace let Salesforce ISVs draw down AWS Enterprise Discount Program commitments to buy adjacent software. EDP burn-down is the reason an unreasonable number of enterprise SaaS contracts get signed at year-end on the AWS Marketplace. It is plumbing, not strategy, and it generated tens of billions in ISV revenue that would otherwise have lived in a separate procurement queue. Office 365 did the same trick for Teams: an item already in the Microsoft enterprise agreement is an item you do not have to defend twice. The SAP–Microsoft alliance announced in 2024 tried to apply the same template to S/4HANA workloads on Azure. The 10 June announcement reaches further: it makes the frontier model itself a draw-down against an existing commitment, not a new SKU bolted on top. The numbers underneath are large in a way that has not registered with the equity market. Oracle's fiscal 2026 guidance — total revenue at least $67 billion, cloud infrastructure up over 70 percent, total cloud up over 40 percent — was already aggressive before the OpenAI integration was made part of the standard catalogue. The Fusion installed base alone runs into tens of thousands of enterprise customers, with DACH-region anchors including Volkswagen, BMW, Allianz, Siemens, Lufthansa and Deutsche Telekom on Fusion ERP or HCM, plus Bosch, Henkel and most of the German Mittelstand on a long tail of E-Business Suite and JD Edwards installations now being migrated to Fusion. Cerner, acquired by Oracle in 2022 for $28 billion, runs the electronic health records of roughly two-thirds of US hospitals and a growing number of European systems; the Codex part of the announcement matters here, because hospital IT teams have been quietly rewriting Cerner integration code for two years. The distribution effect compounds. An Oracle Universal Credit dollar that buys an OpenAI API call is a dollar of OCI gross revenue that did not need a separate sales motion — it converts a procurement department into a passive channel. For OpenAI, whose roughly $122 billion funding round earlier this year was justified in part by the need to push enterprise revenue past the consumer ChatGPT line, this is the cheapest enterprise customer acquisition channel available short of Microsoft. The contradiction is that Oracle has historically been the least loved hyperscaler in the developer community. By making OpenAI the front-door API on OCI, Oracle borrows OpenAI's brand to reach an audience Larry Ellison's company could never reach on its own merits — and OpenAI borrows Oracle's contracts to reach an audience its own sales team cannot scale to without burning ten years of margin.

On 10 June 2026 a16z general partner David Haber, writing under the firm's enterprise banner alongside Martin Casado's category, published “Everything is Recorded Now.” The argument: meeting recording has silently crossed from edge case to default inside AI-native companies, and the accumulated voice corpus — not the CRM, not the wiki — is becoming the enterprise system of record. The reason is operational. Agents onboarded onto a company's recorded meetings outperform those fed only documentation, because culture, edge cases and the unwritten rules live in conversation. For DACH leaders, the essay is less a forecast than a provocation. Germany's Betriebsverfassungsgesetz, Article 88 GDPR and the EU AI Act's 2 August 2026 high-risk obligations make a US-style default-on rollout legally and politically unworkable without a works-council agreement signed first.

The essay opens with a line that has already been screenshotted into a hundred Slack channels: “You should probably assume that everything you say at work is getting recorded from here on out.” Haber, writing for a16z's enterprise practice, is not predicting the shift. He is describing one that, in his view, has already happened inside OpenAI, Shopify and the firm's own portfolio. “OpenAI now runs with essentially everything recorded, with agents standing in for senior leaders in meetings they can't attend,” he writes. The reference example is Granola, the bot-free meeting notetaker that closed a $125 million Series C in March 2026 at a $1.5 billion valuation, led by Index and Kleiner Perkins, on the back of roughly 250 percent quarter-over-quarter revenue growth. Haber calls it out by name: “Granola has better context on a16z's culture, our investments, and how we actually think than almost any other tool we use, because it's been in the room.” The historical reference is Bridgewater. Ray Dalio's policy of taping every meeting, in place since the late 1990s, was for two decades treated as the eccentric house style of a hedge fund obsessed with radical transparency. Dalio's own lawyers initially warned him he was manufacturing discoverable evidence. He bet the opposite way — that recordings would reduce the probability of misconduct and protect the firm in regulatory exchanges — and Bridgewater's unusually thin litigation record became one of the quieter pieces of evidence for the policy. Haber's claim is that what looked like a personality quirk turned out to be infrastructure. The closest analogue from the prior cycle is email becoming archived by default in the early 2000s after Enron and SOX: a behavioural default flipped, and twenty years of e-discovery jurisprudence followed. The narrative pivot is that this is no longer a productivity story; it is a data-moat story. Haber sketches a two-sided ledger. Bottom-up, an AI with full company context becomes a force multiplier for individual contributors who can finally see where the institution is leaking. Top-down, executives get a handle on alignment by sending their agent into the meetings they cannot personally attend. Both sides compound: every recorded meeting makes the model smarter, and the cumulative corpus becomes proprietary in a way a public LLM cannot replicate. “The system of record today is structured data: CRM entries, tickets, docs,” he writes. “But the highest-value context lives in conversation.” The market is responding. Granola is now positioned as an enterprise context layer, not a notetaker. Microsoft has finally pushed Recall out of preview into broader Copilot+ PC availability after two rounds of privacy delays and is integrating it with Purview for enterprise controls. Fireflies, Otter, Sembly and Plaud are each carving a sub-segment: sales calls, platform-native transcription, decision extraction, hardware capture. Glean is buying its way into the conversational layer from the search side. The category that a16z's investment memo from twelve months ago labelled “meeting tools” is being rewritten in real time into “voice-native system of record.”

The thesis breaks on contact with German labour law, and DACH CIOs need to be clear-eyed about why. Recording a work meeting in Germany is not a product-launch decision; it is a co-determination event. Section 87 (1) 6 of the Betriebsverfassungsgesetz gives the Betriebsrat a hard veto over the introduction of any technical device suitable for monitoring employee behaviour or performance. Federal Labour Court (BAG) rulings handed down through 2026 have read “suitable” expansively: a transcription tool does not have to be used for monitoring to trigger co-determination, it merely has to be capable of it. Zoom AI Companion, Teams Copilot, Granola, Fireflies and Otter all meet that bar by construction. The practical consequence is that a DAX40 CIO cannot enable meeting AI by toggling a tenant switch. The rollout requires a Konzernbetriebsvereinbarung negotiated with the central works council that specifies purpose limitation, access rights, retention windows and an explicit ban on performance evaluation use. Layered on top is Article 88 GDPR, which lets member states write their own employment-data rules. Germany's Section 26 BDSG and the Beschäftigtendatengesetz draft circulating in Berlin both require that employee data processing be “necessary” — a sharper test than legitimate interest. Storing every meeting on the off chance an agent might need the context is the opposite of necessity. Fines under Article 88 reach EUR 20 million or four percent of global turnover. From 2 August 2026, the EU AI Act adds a third layer. Annex III, Category 4 designates most workplace AI that scores, classifies or evaluates employees as high-risk, triggering conformity assessment, risk management, human oversight and registration in the EU database. Emotion recognition in the workplace is now categorically banned, which directly limits the “sentiment on the customer call” use case Haber celebrates. For meetings involving European customer data, the high-risk obligations bite again under the December 2027 omnibus timeline. The DAX40 precedent base is unambiguous. Volkswagen's central works council under Daniela Cavallo has spent two years extracting concessions on every AI rollout from Cariad through to HR analytics. Deutsche Telekom and SAP, currently building the Bund's sovereign AI stack for federal administration, have both signed framework AI-Betriebsvereinbarungen that explicitly carve out meeting transcription as a use case requiring its own subsidiary agreement. SAP's own union history — the 2018 fight over the introduction of SuccessFactors performance modules — is the template every German HR director will reach for. The contradiction with Haber's argument is sharp. He writes that “widespread recording simply happens, because it's too hard to stop, and the controls get retrofitted on top.” In Germany, that sequence is illegal. Controls come first or the rollout does not happen. The deeper question for DACH leaders is not whether to record but whether to concede the data moat to US-native competitors who can compound a voice corpus their European rivals are legally barred from building at the same speed. That is the strategic problem the Casado-Haber thesis hands to every German board this quarter, and it does not have a comfortable answer.

Tokenmaxxing is the AI-era practice of measuring progress by how many tokens an organisation consumes from frontier labs. Anthropic, OpenAI and their peers shifted enterprise contracts from per-seat subscriptions to consumption pricing in Q1 2026 and then encouraged customers to chase volume through internal leaderboards, badges and discounted commits. Spend went vertical. Return did not. Return on Tokens, or ROT, is the framework Markie Wagner and Packy McCormick introduced on 10 June 2026 to fix that: value of output minus cost of tokens, divided by cost of tokens, times one hundred. It is the same arithmetic a CFO applies to a new machine or a new hire. For DAX40 finance chiefs entering Q3 reforecast season, ROT supplies the vocabulary they have been quietly missing while signing nine-figure token commits.

A month before publication, Packy McCormick walked into Soho Diner in Manhattan, ordered a milkshake, and sat down across from a founder he had been told about in hushed tones. Markie Wagner, the co-author of the 2023 Founders Fund essay Choose Good Quests, the ex-Waymo ML engineer and Thiel fellow, had gone quiet for two years. McCormick wanted to know why. She asked them to cut up a bowl of fruit, gave him her lore, and then said the line that became the spine of the essay. “She told me, before everyone else came to the same conclusion, that tokenmaxxing was bullshit,” McCormick wrote, “because behind closed doors, the Fortune 500 CEOs she works with were all saying some version of: We committed to all this token spend and I have no idea what we are getting out of it.” On 10 June 2026, Not Boring published the essay to 269,285 subscribers. Wagner used the platform to do two things at once: announce her stealth company Poetic, backed by Founders Fund, Kleiner Perkins, Genius Ventures and OpenAI with a $50M round at roughly $500M, and bury the metric the labs had been pushing on her customers. Her formula is deliberately CFO-shaped: ROT equals value of output minus cost of tokens, divided by cost of tokens, times one hundred. Two levers, value up or cost down. The essay then introduces a companion ratio, Thinking-Doing Ratio, currently sitting near 1000:1 in most AI deployments, which Wagner argues is upside down. The scene Wagner described in Soho Diner had already started leaking into the open. Uber CTO Mihir Sukthankar admitted the company burned through its entire 2026 Claude Code token budget by April, four months in. COO Andrew McDonald told staff Uber was “having trouble justifying the AI spend based on the actual return that one could actually measure.” Uber then capped employees at $1,500 per month for AI tools. Amazon shut down its internal AI leaderboard in May. Legora CTO Jacob Lauritzen told Harry Stebbings the leaderboards “lead to tokenmaxxing, which is people just burn tokens just to look good. That is a really stupid way to do anything.” Ramp's Veeral Patel called the entire regime the “Token Casino: useful software wrapped in mechanics that make spend feel like progress.” Palantir CEO Alex Karp compared tokenmaxxing to “a porn addiction.” Even Sam Altman conceded on CNBC that companies are “spending a ton of money on AI” without knowing when it shows up in revenue, calling it “a huge issue.” McCormick reaches for a historical comparison that lands hard for European industrial CFOs: every cycle has its dumb metric. In the mid-nineteenth century the market rewarded miles of railroad track laid, and roads were duplicated route by route. At the turn of the millennium it was eyeballs. In the 2010s it was top-line gross revenue, which is how WeWork happened. Each metric eventually got disciplined by a return-on-capital framework. Tokens are next. The closest analogue for DACH readers is the FinOps movement that took hold around AWS spend in 2015, when CFOs stopped trusting line-of-business engineers to right-size instances and started instrumenting cloud bills the way procurement instruments steel. ROT is FinOps for inference.

The plumbing is being built around the new vocabulary in real time, and a CFO who knows where to look can already assemble the toolchain. On 4 June 2026, six days before the Wagner essay, Ramp closed a $750M Series F at a $44B valuation, led by ICONIQ, GIC and Ontario Teachers' Pension Plan, with Goldman Sachs Alternatives, D.E. Shaw, Morgan Stanley Investment Management and Generation Investment Management joining. Ramp's pitch deck, according to TechCrunch's reporting, explicitly positions the platform to bring “the same visibility and control to token spending that it already provides for corporate cards, travel, and vendor payments.” CEO Eric Glyman has crossed $1B in annualised revenue and is free-cash-flow positive. Investors did not pay 44x ARR for an expense-card company; they paid it for a token-FinOps platform. The benchmarking layer is moving just as fast. Mercor CEO Brendan Foody told Harry Stebbings on 20VC that Mercor now “spends more on tokens for our internal agents than we are on employee head count,” and predicted that “in five years the average enterprise spends more on compute than headcount.” Mercor's APEX benchmark family scores frontier models on economically valuable professional-services tasks rather than abstract reasoning, and APEX-Agents extends that to long-horizon, cross-application work. APEX is becoming the equivalent of a TPC benchmark for agent workloads: the number a CFO can put next to a token line item to ask whether the model bought was the right model for the job. Gergely Orosz's Pragmatic Engineer Pulse, drawing on conversations with engineers at fifteen companies, reports that enterprise token spend has risen roughly 10x in six months with no sign of slowing, and that GitHub Copilot and Anthropic have begun rate-limiting unprofitable individual users to protect business accounts. The infrastructure for treating tokens as a managed cost category, the equivalent of compute units in a 2015 AWS bill, is now in place. The playbook Wagner sketches and Pragmatic Engineer's field reporting confirms has five concrete levers, and they are the levers that will dominate Q3 2026 reforecast conversations in Frankfurt, Munich and Walldorf. First, per-team caps modelled on Uber's $1,500-per-head ceiling, set against a measured baseline rather than a vendor commit. Second, default-model downgrades, the move Pragmatic Engineer documented at a large SaaS shop that quietly switched its internal IDE default from Claude Opus to Sonnet and saw bills drop without measurable quality loss. Third, multi-model routing, with frontier models reserved for the 20% of tasks Coinbase CEO Brian Armstrong flagged as “IQ-maxing” work and the remaining 80% sent to cheap open-source Chinese models via OpenRouter; the OpenRouter rankings already show this migration in lockstep with consumption pricing. Fourth, prompt-cost coaching, treating verbose chain-of-thought prompts the way procurement once treated unbatched API calls. Fifth, the abolition of internal token leaderboards, the Amazon move, which removes the perverse incentive structure that Veeral Patel calls the Token Casino. The counter-case is not weak. Ed Zitron has spent six months arguing on his Where's Your Ed At newsletter and on Bloomberg that AI “doesn't have ROI, it's nothing like AWS or Uber, and it's got no post-bubble recovery story.” His view: ROT is a polite framework that lets the bubble continue to inflate while the labs run negative 122% non-GAAP margins. Gary Marcus has made a similar argument from a more academic perch. Wagner's essay does not engage Zitron directly, but Poetic's go-to-market answers him by implication, claiming 100x less token usage and “nines of accuracy” at AIG, SoFi and Chime, with AIG CEO Peter Zaffino quoted as saying Poetic has “achieved 99%+ quality outcomes on multi-hour processes, delivering real enterprise value.” One subscriber comment on the Not Boring post sharpened the macro stake: if Wagner is right and tokens are spent only when the world changes rather than every time work happens, lab revenue tracks volatility rather than work volume. The hyperscalers, on that view, end up long volatility, starved in quiet years and fed in chaotic ones. That is a very different business from the one currently priced into Anthropic's reported $350B valuation talk and Microsoft's capex curve.