← W21: Print, Not Pilot W23: Open the Books →

Week 22 · May 25 – May 31, 2026

high signal

The Trillion-Dollar Print

Anthropic closed its Series H May 28 at $65 billion on a $965 billion post-money valuation — overtaking OpenAI as the most valuable private AI company in the world — and shipped Claude Opus 4.8 the same day. Anthropic disclosed a $47 billion annualized gross billing run rate, up from $9 billion at year-end 2025. The same week, Salesforce printed Agentforce ARR at $1.2 billion (+205 percent year over year, +50 percent sequential from $800 million at Q4 FY26); Dell reported $16.1 billion of AI server revenue in a single quarter (+757 percent year over year) with $24.4 billion in AI orders booked; Snowflake disclosed 13,600+ production Cortex AI accounts at +34 percent revenue growth and a $200 million OpenAI partnership; and Cognition raised $1 billion at a $26 billion valuation behind a $492 million Devin ARR (13x year over year). The UK AI Security Institute confirmed two frontier models from different labs can autonomously execute end-to-end 32-step corporate network intrusion simulations — Mythos Preview 73 percent, GPT-5.5 71.4 percent — with offensive cyber capability now doubling every four months. The Trump White House pulled a near-complete federal AI executive order hours before signing on May 21 after opposition from David Sacks, Musk, and Zuckerberg; Newsom signed California's first AI workforce-disruption executive order the same day; the FTC unsealed first criminal charges under the TAKE IT DOWN Act May 20 (140+ named victims) and sent warning letters to 12 platforms. The Mining Software Repositories 2026 conference published the first peer-reviewed quantification of agentic CI failure — AI agents introduce 79 percent of CI failures while performing only 61 percent of fixes across 11,771 GitHub PRs. P-012 ("AI investment boom is a bubble likely to burst by 2026") is one month from target date with hyperscaler 2026 capex tracking $660-690 billion (+77 percent year over year), Q1 venture funding at an all-time record $300 billion, and zero of the five operational burst criteria met — the prediction will resolve refuted at 2026-06-30. The frontier AI valuation print and the AI safety infrastructure withdrawal happened in the same week.

6 predictions updated5 milestones25 companies refreshed

Key Developments

Anthropic closed Series H May 28 at $65 billion on a $965 billion valuation — overtakes OpenAI as the most valuable private AI company — and shipped Claude Opus 4.8 plus Dynamic Workflows the same day

Anthropic closed Series H on May 28, 2026 — $65 billion raised at $965 billion post-money valuation, co-led by Altimeter, Dragoneer, Greenoaks, Sequoia with ~20 co-investors including Blackstone, Brookfield, GIC, Coatue, D1, Temasek, Founders Fund. $15 billion of the $65 billion is previously committed hyperscaler capital (Amazon $5 billion, plus Google TPU, Broadcom, SpaceX GPU agreements). The round overtakes OpenAI's $840 billion February valuation as the most valuable private AI company in the world. Anthropic disclosed an annualized gross billing run rate of $47 billion as of late May, up from $9 billion at year-end 2025, $30 billion in April, and $44 billion by mid-May — an 80x year-over-year leap. OpenAI's CRO previously disputed the April $30 billion figure citing gross-vs-net recognition (~$8 billion overstatement); recognized revenue is likely $33-40 billion on the lower end. Claude Opus 4.8 + Dynamic Workflows shipped the same day — 41 days after Opus 4.7, the fastest premium-tier cadence Anthropic has set. Dynamic Workflows productizes multi-week agentic execution as a first-class primitive, capped at 1,000 parallel subagents per session; Fast Mode is 3x cheaper than prior Opus; the model is 4x less likely to miss coding flaws relative to 4.7. IPO preparation is reportedly active (Wilson Sonsini retained, capital-markets hires complete), targeting a NASDAQ listing summer-to-October 2026. At $965 billion this would be among the largest tech IPOs ever attempted.

Confirms P-001 ↗ AI models will handle most aspects of software eng...Confirms P-021 ↗ 40% of enterprise applications will feature task-s...Challenges P-012 ↘ AI investment boom is a bubble likely to burst by ...Confirms ↗

Counterpoint

The $47 billion figure is annualized gross billing, not GAAP-recognized revenue — and OpenAI's CRO already disputed the April $30 billion claim by roughly $8 billion on gross-vs-net grounds. The actual recognized revenue is likely $33-40 billion on the lower end of that range, which makes the headline "Anthropic overtakes OpenAI on revenue" technically defensible but materially contested. At a $965 billion valuation against $7 billion+ in annual compute spend (mostly through Amazon and Google partnerships that could become competitive), Anthropic is structurally capital-intensive in a way that 2026's most-comparable software companies (Microsoft, Alphabet) are not. Mythos Preview's 77.8 percent SWE-bench Pro and Opus 4.8's "4x fewer missed coding flaws" are Anthropic-reported benchmarks on Anthropic-defined evaluations — Benchmark Theater applies until an independent re-test (METR, NIST, academic) reproduces the numbers. Mythos has already been matched by OpenAI's GPT-5.5 on the AISI cyber chain within weeks — the lead is generational, not lab-specific. The IPO timeline (summer-October 2026) is "speculation plus pre-filings," not a confirmed date. A $965 billion private company with no clean precedent for going public faces SEC disclosure scrutiny that will probably force the gross-vs-net dispute into resolution one way or the other.

Source →

Salesforce Q1 FY27 printed Agentforce ARR at $1.2 billion (+205 percent year over year, +50 percent sequential from $800 million) — 28.6 trillion tokens processed, Slack MCP at one million users in six weeks

Salesforce (CRM) reported Q1 FY2027 on May 27 — Agentforce ARR hit $1.2 billion (+205 percent year over year, +50 percent sequential from $800 million at Q4 FY26). Data 360 + Agentforce combined ARR reached $3.4 billion (+200 percent year over year), including $1.1 billion Informatica Cloud ARR. 3.8 billion Agentic Work Units delivered to date (+111 percent quarter over quarter); 28.6 trillion tokens processed in Q1 alone (+152 percent quarter over quarter). Slack MCP surpassed one million active users within six weeks of GA. 50 percent or more of Agentforce bookings came from existing-installed-base expansion — confirming the land-and-expand thesis is operational at scale. Total Q1 revenue $11.13 billion (+13 percent year over year); FY27 guidance raised to $45.9-46.2 billion. The $1.2 billion Agentforce ARR is the clean inflection — a single named agentic-AI revenue line at SaaS hyperscale, growing 50 percent sequentially in 13 weeks. Combined with Snowflake (13,600+ Cortex AI accounts, $200 million OpenAI partnership), Workday ($500 million agentic ARR from W21), and UnitedHealth (Optum Rx PreCheck cut prior auth approval from 8 hours to under 30 seconds, on pace for 2.5 billion AI transactions), agentic AI is now the earnings-print category, not the pilot-disclosure category, across five different enterprise-software platforms in a single earnings cycle.

Confirms P-021 ↗ 40% of enterprise applications will feature task-s...Confirms ↗ Challenges ↘

Counterpoint

Agentforce at $1.2 billion ARR is roughly 10.8 percent of annualized Q1 revenue ($11.13 billion × 4 = $44.5 billion); growth is hyperbolic but share-of-business is still minority. The 3.8 billion Agentic Work Units and 28.6 trillion tokens are activity metrics, not outcome metrics — AWU measures actions taken, not value delivered, and token volume reflects compute consumed at Salesforce-vendored prices, which is exactly the kind of usage-based revenue line that can be expensive without being valuable. Goldman Sachs's May 2026 productivity note — no meaningful economy-wide AI-productivity relationship — is the macro disconfirming evidence on whether buyers of this $1.2 billion ARR get measurable returns. The 50 percent existing-customer expansion stat is the strongest single counter to the saaspocalypse-eats-Salesforce thesis (it confirms incumbents are pricing per-action AI on top of per-seat contracts at scale), but the renewal-cycle test arrives when net-new ACV needs to absorb seat compression rather than expansion. NRR and churn in Q2 are the only numbers that will tell whether Agentforce is durable revenue or per-action usage that walks at renewal.

Source →

UK AISI confirms two frontier models autonomously complete 32-step corporate cyber chains — Mythos 73 percent, GPT-5.5 71.4 percent — offensive cyber doubling every 4 months, just as the Trump White House pulled the federal AI executive order

The UK AI Security Institute — an independent government evaluation, Tier-A source — published findings in late May that two frontier models from different labs autonomously completed end-to-end 32-step corporate network intrusion simulations. Claude Mythos Preview achieved 73 percent pass rate on expert-level CTFs (3 of 10 runs cleared the full chain); GPT-5.5 matched at 71.4 percent (2 of 10 runs). AISI estimated a human expert requires approximately 20 hours for the same chain. The offensive cyber capability doubling rate is now estimated every four months — down from a seven-month doubling at end-2025. A jailbreak that fully bypassed GPT-5.5 safety guardrails was also found during the evaluation. The same week (May 21), the Trump White House pulled a near-complete federal AI executive order hours before a signing ceremony after opposition from David Sacks ("doomers wanted this"), Elon Musk, and Mark Zuckerberg. Killed provisions included voluntary pre-release model sharing with government (up to 90 days), Treasury-led security vulnerability assessment, and a cybersecurity regime for frontier AI. Trump: "I didn't like certain aspects — it could have been a blocker." Newsom signed California's first-of-its-kind AI workforce-disruption executive order the same day, deliberately contrasting. The capability bar is rising at the exact moment the only US federal pre-deployment evaluation infrastructure being proposed is shelved.

Confirms P-034 ↗ AI capabilities are advancing at a rate that outst...Confirms ↗

Counterpoint

Two model passes at 71-73 percent on a single AISI-designed CTF chain are an alarming data point, but they are not yet a deployment-equivalent capability — both runs were single-attempt evaluations under controlled conditions, not loose-in-the-wild persistent autonomous campaigns. The same week Anthropic published an alignment-research note finding model failures are still dominated by incoherence rather than systematic misalignment — meaning the offensive capability is real but the operational reliability remains low enough to bound it. The Trump EO pull is a policy regression, but Colorado SB 24-205 enforcement begins June 30 and California SB 53 (frontier AI safety plans) is already in force; the EU AI Act Omnibus added new prohibitions on AI CSAM and non-consensual imagery effective December 2026. Forty-plus US states have AI laws enacted or pending in 2026 — the patchwork is not equivalent to federal pre-deployment evaluation but it is not zero. The honest framing — the gap between capability acceleration and evaluation infrastructure is widening, but the absolute level of either is hard to translate into incident probability over a 12-month window.

Source →

What the Evidence Moved

P-001AI models will handle most aspects of software engineering t...

Opus 4.8 + Dynamic Workflows (May 28) productizes multi-week agentic execution as a primitive at 1,000 parallel subagents; GPT-5.5 88.7 percent SWE-bench Verified leads leaderboard. MSR 2026 peer-reviewed data confirms AI fixes failures 4x faster than humans but introduces 79 percent of them.

80%→82%▲ +2pp

P-012AI investment boom is a bubble likely to burst by 2026

Final pre-resolution check, ~30 days to target_date. Anthropic Series H closed $65 billion at $965 billion. Hyperscaler 2026 capex tracking $660-690 billion (+77 percent year over year). Dell Q1 FY27 AI server revenue $16.1 billion (+757 percent year over year). None of the five operational bubble-burst criteria met. Prediction will resolve refuted at 2026-06-30.

5%→5%▼ 0pp

P-02140% of enterprise applications will feature task-specific AI...

Salesforce Q1 FY27 print Agentforce ARR $1.2 billion (+205 percent year over year, +50 percent sequential from $800 million), Snowflake 13,600+ production AI accounts, UnitedHealth Optum Rx PreCheck cut prior auth 8h to under 30s. Five different enterprise platforms shipped agentic ARR or production deployment data in a single earnings cycle.

90%→91%▲ +1pp

P-0332026 is the year everyday people will be fooled by deepfakes

FTC began TAKE IT DOWN Act enforcement May 19; on May 20 first criminal charges unsealed against two individuals (140+ named victims). EU AI Act Omnibus added prohibitions on AI CSAM and non-consensual imagery effective December 2026. Harm side moves from contested to enforcement-confirmed.

70%→72%▲ +2pp

P-034AI capabilities are advancing at a rate that outstrips the e...

UK AISI confirms two frontier models execute end-to-end 32-step corporate cyber chains autonomously (Mythos 73 percent, GPT-5.5 71.4 percent). Offensive cyber doubling every 4 months (was 7 months end-2025). Trump White House pulled federal AI EO May 21 — the only pre-deployment government evaluation mechanism shelved.

Next issue drops Monday

Subscribe to get the briefing before the market opens.

← Print, Not Pilot Open the Books →

Share on LinkedIn Share on X

The Trillion-Dollar Print

Key Developments

Anthropic closed Series H May 28 at $65 billion on a $965 billion valuation — overtakes OpenAI as the most valuable private AI company — and shipped Claude Opus 4.8 plus Dynamic Workflows the same day

Salesforce Q1 FY27 printed Agentforce ARR at $1.2 billion (+205 percent year over year, +50 percent sequential from $800 million) — 28.6 trillion tokens processed, Slack MCP at one million users in six weeks

UK AISI confirms two frontier models autonomously complete 32-step corporate cyber chains — Mythos 73 percent, GPT-5.5 71.4 percent — offensive cyber doubling every 4 months, just as the Trump White House pulled the federal AI executive order

What the Evidence Moved

Company Impact

Anthropic

Salesforce

S&P Global

Tempus AI

Texas Instruments

Snowflake

Southern Company

Visa

Uber Technologies

UnitedHealth Group

Starbucks

Funding Snapshot

Regulatory Calendar

Capability Milestones

Sources