The ROI reckoning arrives — Anthropic laps OpenAI at $44B ARR and turns the corner on margins, but Microsoft kills its Claude licenses and the enterprise token-spend bill comes due

The Threads

The ROI reckoning — Anthropic laps OpenAI, but the enterprise token-spend bill comes due

The bull case and the bear case are now the same sentence: enterprises will pay anything for frontier AI right up until someone asks them to prove it worked. This week both halves printed at once.

The bull half is extraordinary. On the 20VC roundtable, Rory O’Driscoll laid out an Anthropic that “did as much in Q1 as all of last year” — $44B ARR, lapping OpenAI in revenue, with gross margins expanding from 38% to 70% and a $559M operating profit projected for Q2. The trajectory is the tell: gross margin went from roughly negative-60% two years ago, to +34% last year, to 70% now. As O’Driscoll put it, “high growth and improving margins meant profits were inevitable.” This is the same arc Gavin Baker called EBIT-positive in Issue 06 — now with the specific ramp attached. Anthropic is Pareto-dominant: bigger, growing faster, and more profitable than the field.

The bear half landed in the same forty-eight hours, and it is not soft. Microsoft reportedly moved off Anthropic because Opus is too expensive. On All-In, the besties relayed a Fortune-20 CEO who asked for $1B in AI opex savings; six months in, “the team has spent $200 million on tokens with minimal results.” The Uber COO said a full year’s Anthropic credits were burned in four months with gains “probably there but not measurable.” And Kirkland & Ellis is spending ~$500M to roll its own frontier model rather than keep renting one. Two independent enterprise sources, one number that matters: spend is real, measured ROI is not yet.

The structural risk under Anthropic specifically is pricing. “Claude is twice the price of its competitor,” O’Driscoll noted, “and if even half the world says the ROI isn’t there for a premium product, either it has to cut prices to match or it maintains the premium like Apple but cedes share.” The degradation mechanism is mundane and probably inevitable: as enterprises move from experimental to budgeted, a procurement manager mixes a cheaper base model with Sonnet for the hard parts and DeepSeek for the rest. The premium erodes not because Claude got worse but because the buyer got disciplined [forecast: 2026-05-31-001].

The thing to hold is that both can be true for a while. Anthropic’s revenue traction is so strong that, as O’Driscoll said, “it would require a change in the default switches in corporate America for that number to change.” The inertia bet is more of the same. But the trigger to watch is the measurability gate: the moment AI spend eats a visible chunk of the wage bill, the layoff/ROI conversation “becomes quantitative — you have to prove it” before terminating people. When that gate closes, premium pricing is the first casualty [forecast: 2026-05-31-002]. The Onyx Security episode is the quiet confirmation that we’re already past the experimental phase — Maxim Barkogan’s entire business exists because “Anthropic’s revenue is coming from enterprises paying for Claude Code to do work developers used to do,” and those same deployments are now generating incidents real enough to create a security line-item.

The token-flow trade — data-infra and the memory cohort, triple-confirmed

The most actionable idea of the week was also the least dramatic, and it came from a 13-minute CNBC hit with Brad Gerstner. The framing: “data infrastructure companies like Snowflake, Databricks and ClickHouse are all in the token flow. As more tokens are consumed, it accelerates their core data business.” Crucially, he separates them from the AI-application layer — “a very different story than software apps like Salesforce that more directly compete with the AI model companies.” This is the clean inversion of the model-commoditisation worry running through All-In: if models converge and compete on price, the data layer underneath captures the volume regardless of which model wins. Gerstner called the Snowflake quarter two weeks early and it printed a blockbuster — 33% growth, accelerating, versus 27% consensus, full-year guide raised by more than the Q1 beat [forecast: 2026-05-31-003].

But the sharper signal sits one layer down, in memory, and it got three independent confirmations this week. Gerstner: “Micron is up 200% year-to-date” — and he rotated fund capital out of a beloved Snowflake (+10% YTD) precisely because the memory cohort was the better use of dollars. On 20VC, Cerebras CEO Andrew Feldman put a duration on it from the supply side: “memory is the number-two item in the chain after TSMC fab space, and if demand stays high we will continue to see memory shortages for at least the next several years” — prices already up 4-5x in places. And Dan Loeb framed the whole regime: “the SOX is up 40% — I’ve never seen an event like that,” a single Nvidia print three years ago turning semis from “roadkill” into the market’s centre of gravity. When a buy-side legend, a sell-side CEO, and a hedge-fund CNBC regular independently land on the same cohort, the base rate says pay attention [forecast: 2026-05-31-004].

Feldman’s larger argument reframes the entire bubble debate, and it is the strongest counter to this week’s valuation skeptics. “In past bubbles — fiber optics in the late 90s, railroads in the 1880s — the infrastructure buildout was way ahead of demand. AI is the exact opposite. We’re building behind demand. We have a $25 billion backlog. That is not a characteristic of a bubble.” The metering — permitting delays, power constraints, data-centre lead times — is a feature, smoothing demand like on-ramp lights, an analogy he explicitly credits to Gavin Baker’s Watts-bottleneck thesis from Issue 06. The one caveat Feldman plants is for the neo-clouds: “it’s been Nvidia’s strategy to create competitors for the hyperscalers — they’ve funded and over-allocated to the neo-clouds, a dependence which is probably not healthy.” CoreWeave and Nebius bulls should sit with that sentence.

The IPO supply wave — OpenAI + SpaceX + Anthropic vs “100x sales, I wouldn’t buy a share”

The supply side of the AI trade arrived all at once. OpenAI confidentially filed its S1. SpaceX dropped the largest IPO in history. Anthropic is reportedly next. And the most striking thing on the 20VC roundtable was that the skepticism came from the bulls: “100 times trailing sales, my God. I love the S1, but it’s all madness. It’s SolarCity on steroids. It could be the GeoCities deal of the AI era. I wouldn’t buy a share.” When AI optimists are flinching at the multiples, the sentiment signal is worth more than the valuations themselves.

The absorption question is the practical one, and Gerstner framed it as a buyer who has to fund his own participation: “when SpaceX comes, I’ve got to sell something to buy it — we don’t have unlimited capital.” His comfort is depth — “$100 trillion of capital markets relative to $75 billion” — but his actual caution was specific: “the bigger risk is we’ve come a long way quickly. We could have a 10 to 20% consolidation in the SOX — a run-of-the-mill consolidation.” Keep dry powder. Run the 3/6/9 heuristic: if you’re starting from zero, “put 30% to work, then wait for your moments” [forecast: 2026-05-31-005].

The SpaceX-specific reframe is the most investable single idea in the cohort. Gerstner: the launch business and Starlink “alone would not justify the valuation. Him building the data centres that will power all of AI, and training his own AI with the team he acquired from Cursor — that’s the game changer.” And the line of the week: “there’s no human being on the planet better at turning electrons into tokens than Elon Musk.” But Feldman supplies the cold-water footnote: the celebrated OpenAI-Elon compute deal was OpenAI buying “down-rev gear — H100s, B200s, a generation and a half behind. A good deal for Elon, who had them sitting around; not the deal OpenAI wanted.” The same compute story reads as triumph from the SpaceX cap table and as capacity-constrained desperation from a competitor’s chair. Both are in the price now [forecast: 2026-05-31-006]. Nvidia, meanwhile, is “the most profitable company on the planet” (~$200B/yr run-rate) and the stock barely moved on an $81.6B print — because, as O’Driscoll said, “the stock moves on the Delta news, not the total news.” Jensen’s $3-4T-capex-by-2030 call sets the only question that matters for the whole cohort: is the ROI there on the next $2 trillion [forecast: 2026-05-31-007]?

AI’s legitimacy problem — the Pope, the Frankenstein theory, and a 30%-approval technology

The week’s most underpriced risk is not financial — it’s social licence. Pope Leo XIV’s first encyclical, Magnifica Humanitas (235 pages, 42,000 words), is an AI document, warning that “technology takes on the characteristics of those who build, finance and control it” and calling for regulation, a ban on autonomous weapons, and worker-retraining guardrails. Google, Amazon and Meta reportedly lobbied the Vatican to soften it; he refused. On All-In, Sacks agreed the central risk is centralisation of power but argued the likely culprit is government, not a private actor — warning specifically against an “FDA for AI” that could “give notes to model developers” the way trust-and-safety expanded into censorship during the social-media wars.

The sharpest take was Bill Gurley’s, after thirty days reading everything Anthropic has published: “I don’t think they think they’re writing software. I think they’re midwifing a deity.” Built on Dario’s Machines of Loving Grace — “a capitalist economy of AI systems that give out resources to humans based on what the AI systems think makes sense to reward” — plus an 80-page constitution and a “chief philosopher.” Chamath’s overlay is colder and, for an investor, more useful: the doom messaging is partly game theory. “If you want to be unexploitable, get three or four entities in a room, close the door, and dominate them. The referees don’t understand the game.” The halo is real — “poll the intellectual elite on who’s most caring and they’d put Anthropic first.” This is the bear lens on the same Anthropic the first thread valued at $44B ARR: a company whose safety brand is simultaneously sincere and strategically load-bearing.

The investable edge of this thread is the regulatory tail. Sacks sees breadcrumbs toward an open-weight ban — “rhetoric that open-source models lack guardrails, predicate facts being laid to justify an action later.” The paradox he and Jason keep returning to: “China, of all people, is leading the open-weight movement while the US centralises.” A US open-weight ban would idle the cloud infrastructure serving those models and hand the rest of the world a cost-and-customisation edge — a genuine own-goal risk, though probably not a 2026 event [forecast: 2026-05-31-008]. Joe Lonsdale quantifies the social deficit underneath all of it: “America is ~30% positive on AI; China is 75-80%.” His read — AI is inheriting social media’s reputational poison — is why Onyx’s data-sovereignty wedge (“enterprises won’t let Anthropic or OpenAI keep their agent data — they’re data-hungry companies that will train on it”) and Apple’s “intelligence sovereignty” dark-horse positioning are the same trade from different angles: in a low-trust environment, the neutral layer that doesn’t feed the frontier labs captures the anxiety premium.

Software-defined warfare — the commercial-tech curve, China’s fragility, and Palantir+Anthropic live

Two episodes this week, from opposite ends of the political spectrum of guests, converged on the same defense thesis — and it is more investable than it first looks. On Invest Like the Best, a former DoD advisor gave the structural frame: “dictators are enormously strong and enormously weak at the same time — strong because they control the state apparatus, weak because they’re illegitimate and trust no one.” Applied to China: “Xi doesn’t know who’s on our side in his Standing Committee. That is our edge.” He believes the CCP falls eventually the way the USSR did — “it looked strong right up until the last moment; there were Pentagon documents about Soviet strength two weeks before it fell.” This is not a tradeable catalyst, but it’s the directional spine under the entire reshoring-and-defense thesis: a bet on US institutional durability against authoritarian fragility.

The actually-investable insight is the commercial-tech curve. “The evolution of a Ukrainian drone from three years ago to today — 50 iterations, built in garages through commercial supply chains. Things that become permissive and inexpensive in the commercial world have enormously valuable inputs into asymmetric war.” Commercial iteration speed now beats primes’ procurement cycles — which is exactly why Lonsdale’s confirmation lands: “Palantir and Anthropic are being used right now and it’s working really well — one of the most accurately fought wars, least civilian deaths, most bad guys falling fast.” Two independent defense-insider sources, same software-defined-warfare conclusion, direct read-through to the Palantir/Anduril/defense-software cohort [forecast: 2026-05-31-009].

The bottleneck — and therefore the catalyst — is procurement. The DoD advisor was specific: “we need multi-year authorities. One-year money plus four or five continuing resolutions a year, where no new starts are allowed, is killing magazine depth” — naming the Tomahawk and 155mm stockpile gaps directly. For defense-prime and munitions investors, multi-year-procurement legislation is the unlock that turns the deterrence-gap rhetoric into a capex cycle [forecast: 2026-05-31-010]. Sitting above all of it is the one tail risk that would reprice the entire AI-infra supply chain at once: Taiwan. Citing Kevin Rudd, the advisor reads Xi as “a risk-taker who sees retaking Taiwan as the apotheosis of his life’s achievement.” A leader who needs a legacy-defining win to cement an illegitimate mandate is most dangerous near the end of his tenure — and TSMC sits at the centre of that board. Dan Loeb’s two-variable macro — “oil and AI trump everything” — is the bridge: the oil variable runs straight through the Strait of Hormuz and the same geopolitical board the DoD advisor was mapping.

Trade Docket

The memory cohort (Micron / SK Hynix / Samsung) — long, multi-confirmed. Three independent sources this week: Gerstner (Micron +200% YTD, rotated fund capital in), Feldman (multi-year shortage from the supply side, prices +4-5x), Loeb (the semis regime change). Wrong-if: DRAM ASPs roll over for two consecutive months, or a memory-cohort name underperforms the SOX by >15% over 90 days. Watch-trigger: TSMC + memory-maker earnings, HBM contract pricing. Review: 2026-08-31. Why not folded into a thread headline: it’s a continuation of the Issue 06 Gavin Baker DRAM thesis, not a new convergence — but it is the single most actionable line in the issue and must not be buried.

This Week’s Episodes

{/* Episodes are listed in the frontmatter; the site renders them below. */}

Issue 7 is live. 8 podcasts, 7.8 hours of audio, ~18 minutes to read. The reckoning thread is the one to carry forward: watch the measurability gate — the first quarter an enterprise publicly cuts frontier-lab spend on provable-ROI grounds (Microsoft already moved) is the quarter Anthropic’s premium pricing gets tested in the open. Next week: same shows, new threads.