Tech Digest – May 21, 2026
The Capability Frontier
An OpenAI Model Disproves an 80-Year-Old Math Conjecture — With Original Ideas
An internal OpenAI model has disproved Erdős’s 1946 planar unit distance conjecture in discrete geometry, overturning decades of belief that square grids were optimal. The model found an infinite family of point arrangements and built the proof from algebraic number theory rather than geometry — what reviewing mathematician Arul Shankar called “original ingenious ideas” drawn from “a vast array” of mathematical fields. Fields Medal winner Tim Gowers and Princeton mathematician Will Sawin both confirmed and refined the result.
This was a general-purpose model, not a math-specialised system. Less than a year ago, frontier AI was at International Mathematical Olympiad gold level. Epoch AI researcher Yafah Edelman has now pulled her median timeline for solving most Millennium Prize Problems forward to 2032.
Note: The speed of this progression reshapes planning assumptions. If the gap between “gold-medal math student” and “disproving open conjectures with original ideas” closed in under twelve months, institutional timelines built on “AI can’t do X yet” are depreciating faster than the plans that depend on them.
Sources: OpenAI, Arul Shankar (remarks), Sam Altman, Noam Brown, Yafah Edelman
Governing the Frontier
White House Postpones Frontier AI Executive Order as OpenAI Pursues State-Level Strategy
The White House has postponed an executive order that would have established a voluntary 90-day pre-release review for frontier AI models — treating new releases somewhat like regulatory submissions. President Trump said he “didn’t like certain aspects” of the draft, which had been negotiated with major AI labs including OpenAI and Anthropic. The review window itself was contested, with some companies pushing for 14 days instead of 90.
Meanwhile, OpenAI’s top lobbyist is pursuing what Politico describes as a “reverse federalism” strategy — shaping state-level AI legislation the industry can accept, as federal regulation stalls.
Note: The EU’s AI Act is already in force. The US is still debating whether voluntary disclosure should last two weeks or three months. For EU institutions, this widening transatlantic governance gap shapes which products and services will be compliant and deployable in their jurisdictions first — and which vendors will treat European regulatory requirements as a competitive advantage rather than a burden.
Sources: The Information, CNN, Axios, Politico
Silicon, Power, and Zoning
Nvidia Posts Record $81.6 Billion Quarter — Then Concedes China to Huawei
Nvidia reported $81.6 billion in Q1 revenue, up 85% year-over-year, with data centre revenue alone reaching $75.2 billion — up 92% from a year ago. The company announced an $80 billion share buyback and increased its quarterly dividend 25-fold to $0.25 per share. Revenue guidance for Q2 is $91 billion.
But the quarter’s other headline came from Jensen Huang acknowledging that Nvidia has “largely conceded” the Chinese AI chip market to Huawei — the clearest admission yet that export controls have permanently bifurcated the global chip supply.
Note: $75 billion in quarterly data centre revenue makes Nvidia’s infrastructure business larger than most countries’ annual IT budgets. The China concession formalises what de-risking looks like in practice: two parallel chip ecosystems, with procurement strategies increasingly defined by which side of the divide you operate on.
Sources: Nvidia Newsroom, CNBC
$15 Billion a Year for Compute, $2.8 Billion in Turbines, and One Missouri Town Says No
Anthropic is now paying SpaceX’s xAI division $1.25 billion per month for compute access — $15 billion annually — and is scaling onto Nvidia GB200 capacity in xAI’s Colossus 2 facility through June. The arrangement places Anthropic in the surreal position of bankrolling its competitor’s infrastructure landlord. Separately, xAI is purchasing another $2.8 billion in turbines for data centre power, even as it faces lawsuits over existing generators.
The demand is outrunning manufacturing: Seagate’s CEO conceded that building new factories would simply “take too long” relative to AI demand. And the physical footprint is generating resistance — St. Charles City, Missouri, voted this week to effectively ban large-scale data centres, a reminder that zoning boards sit between ambition and infrastructure.
Note: When a single AI lab’s annual compute bill exceeds the GDP of 80 countries, the bottleneck isn’t algorithms — it’s concrete, copper, and community consent.
Sources: Axios, Tom Brown, TechCrunch, CNBC (Seagate), Fox2Now
Compute Leaves the Planet
SpaceX Files for IPO Claiming a $28.5 Trillion Market — Bezos Agrees Orbital Data Centres Are “Very Realistic”
SpaceX has filed its IPO prospectus on Nasdaq (ticker: SPCX), targeting up to a $2 trillion valuation and claiming a $28.5 trillion total addressable market — roughly the entire US GDP. Of that, $26.5 trillion is attributed to AI, spanning infrastructure, enterprise applications, consumer subscriptions, and a Tesla-collaborated AI agent platform called Macrohard. The filing reveals Q1 2026 losses of $4.28 billion on $4.69 billion in revenue.
Jeff Bezos, whose Blue Origin competes with SpaceX, called space-based data centres “very realistic” but Musk’s 2-3 year timeline “a little ambitious.” The debate over orbital compute has quietly collapsed from physics to scheduling.
Note: Two individuals with a combined net worth exceeding $500 billion agree that data centres belong in orbit and argue only about when. Whether the timeline is two years or ten, any institution investing in large-scale digital infrastructure today should consider what happens to terrestrial facility economics when compute starts migrating upward.
Sources: CNBC (SpaceX IPO), Fortune, CNBC (Bezos)
The AI Business Comes of Age
Anthropic Projects First Operating Profit at $10.9 Billion in Quarterly Revenue — OpenAI Prepares IPO Filing
Anthropic has told investors it expects $10.9 billion in Q2 revenue — roughly 130% sequential growth from Q1’s $4.8 billion — and approximately $559 million in operating profit, its first ever. Last summer, the company projected no full-year profit until at least 2028. The acceleration was driven by compute cost ratios improving from 71 cents per revenue dollar to a projected 56 cents.
OpenAI is preparing to confidentially file for an IPO as soon as this week, with Goldman Sachs and Morgan Stanley as lead underwriters. The company, last valued at $852 billion, targets a September listing — just days after defeating Elon Musk’s legal challenge to its for-profit conversion.
Note: Two years ago, these were research labs burning through venture capital. Now one is profitable and the other is going public. The transition from speculative AI venture to established enterprise supplier changes the procurement conversation — these are no longer experimental vendors with uncertain futures.
Sources: WSJ (Anthropic), CNBC (Anthropic), WSJ (OpenAI IPO), Bloomberg
Displacement Signals
Intuit Cuts 3,000 Jobs — 17% of Its Workforce
Intuit, the maker of TurboTax and QuickBooks, is cutting approximately 3,000 employees — 17% of its global workforce. The restructuring will trigger $300-340 million in charges. CEO Sasan Goodarzi’s internal memo cited AI-focused simplification, though he later told CNBC the cuts had “nothing to do with AI.” Revenue grew 10% to $8.56 billion in the quarter, its slowest expansion rate since 2024.
Note: The memo says AI. The interview says not AI. This is the pattern to watch: companies restructure around automation capabilities while publicly disclaiming the connection. For workforce planners, the stated reason matters less than the structural outcome — 3,000 fewer positions at a company whose core products are being rebuilt with machine learning.
Sources: Reuters, TechCrunch, CNBC
Commonwealth Prize Winner Flagged as AI-Generated — Authorship Enters the Forensic Era
The winner of the Caribbean category of the 2026 Commonwealth Short Story Prize was flagged as 100% AI-generated by detection tool Pangram — a result independently confirmed by WIRED and Anthropic’s Claude. A second regional winner was also flagged. The Commonwealth Foundation acknowledged it did not use AI detection during judging, citing concerns about artistic consent for unpublished work.
Note: If a prize jury, a detection tool, and an AI model all reach different conclusions about whether a human wrote something, every institution that procures, evaluates, or certifies written work faces the same question. Authorship is no longer a fact — it’s a forensic assessment.
Sources: The Guardian, The Free Press
Today’s items share a single thread: the distance between “experimental” and “established” is collapsing across every dimension simultaneously. A general-purpose AI disproves open mathematics conjectures with original ideas. An AI lab posts its first quarterly profit while its largest competitor prepares to go public. The two wealthiest people on Earth agree data centres belong in orbit and argue only about when. Meanwhile, the US government can’t agree on how long to review a model before release, a literary prize jury can’t tell whether a human wrote the winning entry, and a financial software company cuts 17% of its workforce while its CEO denies AI had anything to do with it. The institutions that will navigate this well aren’t the ones waiting for certainty — they’re the ones building decision frameworks that assume the ground is still moving.
