Last updated on July 6th, 2026 at 07:50 am

Frontier AI’s most durable product is the year after next. This month the pitch ships with a safety framework, two confidential S-1s, a trillion-dollar reserve price, and a 5 percent offer to the referee.

Sam Altman published an opinion piece in the Financial Times on July 1 proposing a safety framework for artificial intelligence. He calls it simple. Safety standards, he writes, are a prerequisite to the broad distribution of AI, and within another year or two the industry expects to have built systems of “astonishing power,” warranting an American-led international forum modeled on the IAEA to decide who gets access to them and on what terms. The following day the same paper reported, citing two people familiar with the talks, that OpenAI has proposed handing the US government a 5 percent stake in the company.

The venue is new. The promise is eleven years old, and it has arrived, once again, in the exact season the money needs it.

Anthropic raised in late May at a $965 billion valuation and filed a confidential S-1 on June 1. OpenAI filed its own confidential S-1 in early June. SpaceX went public on June 12 and slid from above $225 to the low $150s within two weeks. By June 26, the New York Times was reporting that OpenAI’s advisers had counselled delay and that the company was leaning toward a 2027 listing rather than accept a valuation below the $1 trillion Altman has set as his floor. SoftBank, OpenAI’s second-largest outside shareholder, fell as much as 13 percent on the news, roughly $38 billion in market value gone in a session. The FT op-ed landed four days later, two weeks after a G7 summit where the administration and AI executives discussed a US-led approach to global AI rules.

In IPO season, prospectus prose migrates onto the op-ed page. These announcements are the season’s weather.

Eleven Years of Imminence

The “astonishing systems” have a publication history.

• 2015. Altman’s blog calls superhuman machine intelligence “probably the greatest threat to the continued existence of humanity.”
• 2021. Moore’s Law for Everything proposes an American Equity Fund: large companies contribute 2.5 percent of their market value in shares, distributed to citizens.
• January 2023. The downside case, he tells an interviewer, is lights out for everyone.
• February 2023. Planning for AGI and Beyond commits OpenAI to supervising its own careful transition.
• May 2023. Senate testimony, followed weeks later by his signature on a statement ranking AI extinction risk alongside pandemics and nuclear war.
• March 2024. He tells The Advocate the biggest misconception in AI is confusion over whether it is a tool or a creature, and that ChatGPT is “clearly a tool.” Creature talk, he says, belongs to science fiction.
• September 2024. Superintelligence may be “a few thousand days” away. OpenAI closes a $6.6 billion round days later.
• December 2024. AGI will arrive sooner than people expect and matter less than they fear. Reporting has by then surfaced a contractual definition of AGI pegged to roughly $100 billion in profits.
• January 2025. OpenAI is now confident it knows how to build AGI, and agents may “join the workforce” within the year. A record SoftBank-led round follows within the quarter.
• February 2025. The cost of using a given level of AI, he writes, falls roughly tenfold every twelve months.
• June 2025. The takeoff has started. 2026 will likely bring systems that produce novel insights; 2027, robots that do useful work in the physical world.
• December 2025. The warnings pivot to frontier dangers: cyber offense, biosecurity, self-improving systems.
• April 2026. A 13-page policy paper invokes the New Deal, proposes a public wealth fund to hold stakes in AI companies, and warns of a possible world-shaking cyberattack within the year. It publishes the same day as an 18-month New Yorker investigation into the distance between OpenAI’s safety commitments and its safety spending.
• July 2026. Astonishing power within a year or two, standards as the precondition for distribution, an American-led forum to administer both. The next day, the FT reports OpenAI’s proposal to hand Washington 5 percent of the company.

The 2025 window closed on agents that remain atomized, heavily scaffolded, and error-prone without human oversight. The 2026 window for novel-insight systems is half spent. The op-ed opens a fresh one, and the fresh one expires conveniently after a 2027 listing.

The Question Is the Product

Altman asked: is the machine is a creature. Raised by the chief executive of the company selling the machine, the question does the marketing and the answer does the liability management. A tool attracts product law, consumer protection, and warranty claims. A near-creature of astonishing power attracts an international agency. The FT framework requires the second framing, and eleven years of oscillation have kept it available. And the questions have migrated. Is it a creature? Is it AGI? Is it conscious? Is it dangerous? the pattern of the questions reflects the arc of the market(s).

The oscillation follows the money. The register inflates when capital forms: a few-thousand-days horizon ahead of a $6.6 billion close, declared confidence ahead of the SoftBank round, takeoff language into the restructuring year, astonishing power into the S-1 window. It deflates when accountability approaches: a tool when job anxiety peaks, an AGI that matters less once its contractual definition surfaces. The vocabulary carries strike prices. Microsoft’s agreement reportedly defines AGI in the neighborhood of $100 billion in profits. Thirty-five billion dollars of Amazon’s commitment reportedly unlocks when OpenAI goes public or reaches AGI, whichever comes first.

When a word has a strike price, read its usage as an option, and read the warnings the same way. A warning about the danger of your own product class is a capability claim wearing a safety costume, and it manufactures demand for the gatekeeping architecture its author then volunteers to design. This series has called that pattern Scare Theatre. The FT piece is its multilateral edition.

The Substrate Holds Still

Underneath every frontier system in production sits the architecture published in 2017: an autoregressive transformer sampling the next token from a probability distribution. The releases of the past two years changed what surrounds that substrate. Reinforcement learning after pretraining, test-time compute, tool scaffolds, retrieval, memory bolted on from outside. Engineering around the architecture has been real and sometimes impressive. Evolution of the architecture has been absent, and the limits that follow from it have held.

December 2025 ran the natural experiment. Vendors expanded context windows and shipped cross-session memory within weeks of each other, and the failure arrived with the feature: conflation across sessions, topic bleed across sessions, injected recollections the model could neither rank nor quarantine. The failure was specified before the feature shipped. Without strict and revocable semantic dominance, the two primitives current architectures lack, text inserted into context arrives with unbounded authority. Memory systems built on that substrate mitigate by instruction and etiquette. The architecture offers them nothing stronger.

The same holds for the pattern Evans’ Law describes: the longer a model reasons, the greater the likelihood that a response will be incorrect, until the likelihood of an incorrect answer exceeds the likelihood of a correct one. It is a statement about predictability, and it has held across model generations because nothing in the generation cycle touches its cause. Independent research has since traced the same shape: long-context performance degrades unevenly and degrades faster amid plausible distractors, and Apple’s 2025 work found reasoning models collapsing outright past complexity thresholds.

The vendors’ own instruments agree. OpenAI’s system card showed o3 hallucinating at roughly double the rate of o1 on a benchmark of facts about people, with o4-mini worse still. Its September 2025 research paper conceded the structural point: models guess because training and evaluation reward confident guessing over abstention. Proper nouns, scientific nomenclature, and technical language outside code stagnated for the same reason, and it is a reason, a mechanism rather than a mood.

The improvement boundary follows the verifiability boundary.

Post-training reinforcement inflects wherever a reward can be checked. Code compiles or it does not. Math verifies or it does not. Probability mass migrates to the checkable, and everything on the unverifiable side of the line, which is most of language and nearly all of judgment, holds still or degrades as the mass moves. Meta’s founding chief AI scientist reached the corresponding conclusion from inside the largest budget in the industry: Yann LeCun left in late 2025 to build world models, saying plainly that scaled language models will fall short of the destination. Absent architectural evolution, persistent state, verifiable memory, bounded authority over context, the mechanism for the promised leap is missing from the machine being sold.

Code Maps the Boundary

Coding is the counterexample everyone reaches for, and it proves the rule. Code improves because code is the one large domain humans built for machine verification: grammars constrain output, compilers reject error, tests execute truth. It is also a bounded use case. Software is an enormous market and a thin slice of language, and progress inside the slice says nothing about the terrain outside it, where the December failures live.

Scoreboards the Players Own

Leaderboards are not usage.

The Leaderboard Illusion documented Meta privately testing 27 variants of Llama 4 on Chatbot Arena and publishing the winner, inside a system that grants the largest labs asymmetric sampling and data access. The Llama 4 that charmed the Arena differed from the Llama 4 that shipped. FrontierMath, built to resist gaming, turned out to be funded by OpenAI, which held access to much of the problem set, a fact disclosed after the o3 headline score. An instrument funded, populated, and headlined by the contestants measures marketing.

Usage measures renewal under production mess, and the two diverge exactly where capability diverges from reliability. Even the strongest pro-progress dataset carries the split in its fine print: METR’s task-horizon curve doubles on schedule at a 50 percent success bar, collapses to roughly a quarter of the length at 80 percent, and drops further on tasks its own raters flagged as messy. Production is the messy 80 percent column. The chart on the slide is the 50 percent one.

The Invoice Arrives Inverted

In February 2025 Altman wrote that the cost of a given level of AI falls about tenfold a year. The invoices ran the other direction. A 2026 analysis of benchmark economics found that while per-token list prices fell, the cost of frontier-level performance rose approximately exponentially, around eighteenfold per year, because marginal gains at the frontier demand disproportionate inference; on one flagship science benchmark, roughly half of measured progress tracked rising inference spend rather than price-independent advances. Gartner puts agentic workflows at five to thirty times the tokens per task of a chatbot. The FinOps Foundation found 73 percent of enterprises over their AI cost projections. And the list prices stopped cooperating too: OpenAI and Anthropic have raised prices and imposed rate limits under surging demand, and unlimited tiers are disappearing across the market.

The advertised curve points down. The paid curve points up. That is the inversion this series described when it laid out the inverted bubble thesis, and the ledger at the top of the stack confirms where the losses pooled: audited documents reviewed by the Financial Times show OpenAI generated $13.07 billion in 2025 revenue against $34 billion in costs, a $20.9 billion operating loss.

Meta is running the same math at hyperscale. In April it raised 2026 capital expenditure guidance to between $125 billion and $145 billion, nearly double the $72.2 billion it spent in 2025 and more than 2024 and 2025 combined, citing higher component prices among the drivers. Multi-year infrastructure commitments reportedly grew by $107 billion in a single quarter, and an executive conceded the company keeps underestimating its compute needs. The stock fell on the guidance while Alphabet and Amazon rose the same week on demonstrated AI revenue.

The spending now anchors a mortality debate. Julia Angwin’s May op-ed in the New York Times, “Meta Is Dying. It’s About Time,” argued the company has the “stink of death” on it: the first decline in daily active users since Meta began reporting the figure, $59 billion in long-term debt after doubling in a year, a $27 billion Louisiana data center financed off the books, ad load and pricing cranked hard enough to lift revenue per user 27 percent in a quarter, a bellwether courtroom loss on addictive design with six figures of similar claims queued behind it, all of it following $80 billion sunk into the metaverse. A Wall Street Journal columnist called the spending growth increasingly unsustainable. The obituary overshoots as finance, since the company still books $56 billion a quarter at a 41 percent operating margin. It lands as diagnosis. Meta’s founding chief AI scientist left saying this architecture will fall short, and the company answered with $145 billion for more of it.

The Efficient Alternative

While the American stack raises prices, China ships efficiency. DeepSeek released V4 in late April: open weights, a million-token context window delivered at a fraction of V3’s inference cost through a hybrid attention design, trained and served on Huawei Ascend and Cambricon silicon rather than Nvidia hardware. Its Pro tier lists at $3.48 per million output tokens against roughly $30 at OpenAI and $25 at Anthropic for comparable work; its Flash tier lists at 28 cents. Stanford’s 2026 AI Index concluded Chinese labs have effectively closed the performance gap, and Chinese open-weight models led by Qwen, MiniMax, and DeepSeek now carry about a third of global LLM usage, up from nearly nothing in late 2024. Kai-Fu Lee calls the resulting labs “skinny athletes,” and the description travels: constraint bred the efficiency that now prices against the constraint’s authors.

Export control did the breeding. Denied top-end chips, Chinese labs innovated on sparse attention, FP8 training, and mixture-of-experts routing, and then pointed the savings at the invoice problem American vendors created. Nvidia’s own chief executive has warned publicly about the day Chinese frontier models ship on Huawei silicon first. This series documented the first act: the Mythos and Fable suspension pushed American enterprise demand toward the Chinese open-weight stack, and the reversal arrived too late to reverse the migration. Rising US token costs now push the price-sensitive remainder in the same direction. The FT framework proposes to formalize the gate at the moment the economics dissolve the fence.

The Referee Takes a Position

The safety framework was half the filing. The day after the op-ed, the FT reported, citing two people familiar with the talks, that Altman and other OpenAI executives have proposed that leading US AI developers each allot 5 percent of their equity to a government vehicle modeled on the Alaska Permanent Fund, the state corporation seeded with oil revenue that pays residents an annual dividend and supports Alaska’s budget. OpenAI’s 5 percent would be worth roughly $42.6 billion at its March valuation. Altman has discussed the idea with Trump, Commerce Secretary Howard Lutnick, and Treasury Secretary Scott Bessent, and met with Bernie Sanders in recent weeks. The talks have reportedly run for more than a year, Bloomberg’s account has the arrangement potentially extending to Anthropic, Google, and Meta, and the participation of all three is unconfirmed. Anthropic has separately floated a tax-funded “digital dividend.”

The redistribution instinct is the oldest continuous thread in the corpus. Moore’s Law for Everything proposed the citizen-equity structure in 2021. Altman funded the largest basic income study in the United States through OpenResearch. The April policy paper proposed the public wealth fund a stake like this would presumably seed. Whatever else the offer is, it is long-considered. It is also precisely placed: pitched to the administration beginning in early 2025, surfaced publicly in the week of the S-1 delay coverage and the safety op-ed, weeks after executive orders asked companies to volunteer pre-release model access to Washington for up to 30 days and directed federal agencies to accelerate adoption of the most advanced systems. Altman praised the access order publicly.

The dividend framing invites doubt. The Alaska fund pays citizens and props a budget, and this proposal names no distribution machinery at all. What the administration’s equity practice does name is leverage: reported stakes in Intel, IBM, and a run of quantum and critical-minerals firms, acquired as instruments of industrial policy, with a sovereign wealth fund ordered into existence in February 2025 and waiting for assets. The 2008 Treasury positions in banks and automakers were emergency exits. These are entrances. On that record, a 5 percent tranche of the AI sector reads as the state buying alignment and the companies buying protection, with the basic income language supplying the ribbon.

The Altman op-ed proposes Washington as the standard-setter deciding which systems reach broad distribution and on what terms. The stake makes Washington a shareholder in every company the standard privileges. A government whose fund appreciates with the two-year promise acquires a fiscal interest in the promise being believed, and in the pricing power the access regime protects. The uniform 5 percent draws the national-champion perimeter in equity: inside the fund, inside the forum, inside the standard. For buyers outside the United States, exposure this series has mapped as jurisdictional, through the CLOUD Act and export directives, gains a fiduciary layer. Procuring the American stack now means procuring from vendors the US Treasury is positioned to partly own.

The Case for the Curve

Inarguably, ask horizons on METR’s 50 percent measure have genuinely doubled on a steady cadence. Coding agents changed real workflows. Scaffolding is how every platform matured; compilers civilized assembly, and harnesses may yet civilize this. On China, the Council on Foreign Relations reads V4 as trailing the closed frontier by roughly half a year, short of chips to serve its best tier at scale, with subsidy and distillation questions attached. Per-token list prices for fixed capability levels did collapse. And the warnings may be sincerely held. So may the redistribution: the basic income advocacy runs back a decade, through a funded study and a published blueprint, and the dividend instinct could be entirely genuine.

All of it can be true at once, alongside reliability horizons a quarter the advertised length, gains pooled inside the verifiable slice, frontier costs compounding eighteenfold a year, and a flagship system that has stayed one to two years away since 2015. Sincerity and function are separate questions. The incentive structure selects which sincere beliefs reach the op-ed page, and in which month.

What the Framework Is For

Read as governance, the FT proposal is thin: a forum, a comparison to the IAEA, a promise of standards to be specified later. Read as positioning, it is precise. Standards as a prerequisite to distribution converts a capability narrative into allocation authority, administered from Washington, proposed two weeks after a G7 where that arrangement was the agenda. The equity proposal completes the circuit: the standard-setter becomes a shareholder in every firm the standard privileges, and the state sits on both sides of the transaction it referees. It arrives as the closed American stack raises prices, as the industry’s second-biggest spender posts its first user decline against a $145 billion compute bill, and as an open alternative undercuts the incumbents by a factor of ten to a hundred. The framework asks the world to treat the two-year promise as load enough to bear all of that.

The systems of astonishing power remain a year or two away. They have been a year or two away since 2015. The valuations cannot wait that long, and the announcements know it.

Jen Evans is Principal of Pattern Pulse AI and co-founder of Tech Reset Canada.

References:

• FT op-ed (July 1); Altman blogs (2015, Feb 2023, Sept 2024, Jan 2025, Feb 2025, June 2025); StrictlyVC interview; CAIS statement; The Advocate interview; The Information on the Microsoft AGI definition; Forbes on the Amazon IPO-or-AGI unlock; NYT June 26 IPO-delay report; New Yorker investigation (April 6); The Leaderboard Illusion (arXiv, April 2025); Epoch/FrontierMath disclosure; OpenAI o3/o4-mini system card and September 2025 hallucination paper; Apple reasoning paper (June 2025); METR task-horizon work; arXiv “Price of Progress” (frontier cost ~18x/yr); Gartner March 2026 agentic token analysis; FinOps Foundation 2026 report; Fortune (April 24) on OpenAI/Anthropic price hikes and DeepSeek V4 pricing; Meta Q1 2026 earnings; Angwin NYT op-ed (May 8); WSJ Heard on the Street (Fitch); Stanford AI Index 2026; Fortune (June 16) on one-third usage share and Kai-Fu Lee; CFR on V4; LeCun departure coverage
• 
  Prior B2NN pieces: Scare Theatre, Inverted Bubble, OpenRouter token-share, Mythos/Fable sequence, Two Missing Primitives, Evans’ Law paper.