The Big Picture
Today’s headlines crystallize a strategic pivot both frontier labs made over the last 10 days: a model company alone doesn’t capture the enterprise dollar — the deployment layer does. OpenAI’s $4B Deployment Company, formally launched this morning with 19 PE/consulting partners (TPG leads, Advent / Bain Capital / Brookfield as co-leads) plus the Tomoro acquisition for ~150 forward-deployed engineers, is the explicit OpenAI mirror to Anthropic’s $1.5B Blackstone / H&F / Goldman JV from May 4. Both labs are standing up their own Accenture-killers, with majority equity control and PE money pre-committed, billing into the same enterprise budget Accenture, Deloitte, McKinsey, BCG, and Infosys have owned for two decades. Same morning: Cerebras formally files the amended S-1/A locking in Sunday’s $150–$160 leak — $4.8B raise, $48.8B fully-diluted valuation, ~28% over Friday’s midpoint — pricing Wednesday, trading Thursday under CBRS. Nadella tells the Oakland jury Microsoft’s $13B was never a gift and Musk never raised concerns to him directly. Claude Code v2.1.139 ships an agent dashboard for tracking parallel sessions — the Software 3.0 plumbing finally getting concrete. The EU pressures both labs for cyber-model access; OpenAI says yes to GPT-5.5-Cyber, Anthropic still says no to Mythos. Thursday is load-bearing: Cerebras debuts, trial closing arguments, Beijing summit opens.
Evening update: Sutskever testified Monday afternoon with the day’s most explosive disclosures: his OpenAI stake is worth ~$7B, he spent about a year gathering evidence of Altman’s “consistent pattern of lying” at the board’s request, and the remaining board met with Anthropic post-ouster about a merger where Anthropic would take over OpenAI. Nadella’s morning testimony delivered the colorful additions — calling November 2023 “amateur city” and a Kevin Scott text reading “I don’t want to be IBM and OpenAI to be Microsoft.” Bret Taylor followed; Altman testimony pushed to Tuesday morning. Separately, OpenAI launched Daybreak this afternoon — the full enterprise cyber play built on GPT-5.5-Cyber + Codex Security with 21 named partners (Cloudflare, CrowdStrike, Palo Alto, Akamai, Trail of Bits, SentinelOne, Okta, plus more) — converting the morning’s EU-access maneuver into the explicit Glasswing / Mythos competitor.
OpenAI officially launches the OpenAI Deployment Company — $4B initial capital, 19 PE / consulting / SI partners (TPG leads with Advent, Bain Capital, Brookfield as co-lead founding partners; BBVA, B Capital, Emergence, Goldman Sachs, SoftBank Corp, Warburg Pincus, WCAS also founding), majority OpenAI-owned and -controlled; Axios puts the implied valuation at $14B; OpenAI simultaneously acquires Edinburgh-based Tomoro AI for the Deployment Company’s ~150 forward-deployed engineers (Tomoro founded 2023, clients Mattel / Red Bull / Tesco / Virgin Atlantic)
Today OpenAI, Axios, Bloomberg, PYMNTS, Tech Startups, Constellation Research, BBVA, Tomoro
OpenAI formally launched the OpenAI Deployment Company this morning, the entity previewed as “The Deployment Company” in The Information / FT reports a week ago. The vehicle is majority-owned and controlled by OpenAI, capitalized with $4B at launch from 19 named investment firms, consultancies, and system integrators. TPG leads with Advent, Bain Capital, and Brookfield as co-lead founding partners; BBVA, B Capital, Emergence Capital, Goanna, Goldman Sachs, SoftBank Corp, Warburg Pincus, and WCAS are also founding partners. Axios puts the implied valuation at $14B. In a coordinated announcement, OpenAI agreed to acquire Edinburgh-based Tomoro AI, a 2023-founded enterprise AI consulting firm with about 150 forward-deployed engineers and clients including Mattel, Red Bull, Tesco, and Virgin Atlantic — Tomoro becomes the Deployment Company’s Day-One engineering bench. The company’s explicit mandate: embed OpenAI Forward Deployed Engineers inside Fortune 500 client orgs to deploy frontier AI to production workflows.
Per the Axios / Bloomberg framing, this is the OpenAI mirror to Anthropic’s $1.5B Blackstone / Hellman & Friedman / Goldman Sachs JV announced May 4 — two frontier labs both standing up captive PE-backed deployment arms within seven days of each other means the strategy is now the playbook, not the experiment. Tomoro’s ~150 engineers gives OpenAI a deployment bench Anthropic is still recruiting for. The $14B implied valuation also means OpenAI just majority-owns a freshly-marked services entity worth more than most mid-cap public consulting firms — pre-IPO yield product wearing JV clothes, billing into the same enterprise budget that has bankrolled Accenture, Deloitte, McKinsey, BCG, and Infosys for two decades.
Cerebras formally files the amended S-1/A this morning — range officially $150–$160 / share (up from $125–$135 raised Friday, up from $115–$125 May 4); 30M shares marketed for up to $4.8B in proceeds; $48.8B fully-diluted valuation (vs. $23B at the February funding round); pricing target Wednesday May 13, trading begins Thursday May 14 on Nasdaq Global Select Market under CBRS; orderbook still running 20x+ oversubscribed at the new band
Today CNBC, Stocktwits, GuruFocus, Investing.com, Parameter, TradingKey
Cerebras filed the amended S-1/A this morning that formalizes Sunday evening’s Reuters leak: price range officially $150–$160 per share, up from the $125–$135 band lifted to Friday afternoon and the original $115–$125 in the May 4 filing. Shares marketed climb to 30M from the original 28M. At the top of the new range, the IPO raises roughly $4.8B in proceeds — up from ~$3.5B under the original terms. The company would be worth up to $48.8B on a fully diluted basis, more than double the $23B mark from February’s funding round. CNBC confirms pricing is targeted for Wednesday May 13 with the stock to begin trading Thursday May 14 on the Nasdaq Global Select Market under ticker CBRS. The orderbook is still running 20x+ oversubscribed at the higher band per Reuters / CNBC sources, with Morgan Stanley, Citigroup, Barclays, and UBS leading.
Per CNBC, $48.8B fully-diluted is the cleanest number yet for the public-market half of the compute trade — up from $23B in February and ~28% above Friday’s midpoint after two upward price-range revisions in three trading sessions. The orderbook running 20x oversubscribed at the lifted $125–$135 band is what forced the second raise inside three sessions; institutional demand moved faster than the underwriters could re-price. The structural read: in the 96 hours after Anthropic’s $200B GCP / $1.8B Akamai / $50B-round-at-$900B triple drop, the public market discovered it was under-priced on AI-chip pure-play exposure by roughly a third. Thursday’s CBRS print sets the comparable that every Anthropic-and-OpenAI valuation conversation through year-end will reference.
Claude Code v2.1.139 ships Monday morning — adds claude agents Research Preview (single dashboard listing every Claude Code session, running / blocked-on-you / done), new /goal command (set a completion condition, Claude keeps working across turns with live elapsed-time / token overlay; works in interactive, -p, and Remote Control), /scroll-speed tuner, claude plugin details, transcript-view navigation (?, {, }, v), hook args: string[] exec form and continueOnBlock, MCP stdio servers now receive CLAUDE_PROJECT_DIR, subagent OTEL spans carry agent_id / parent_agent_id
Today Claude Code changelog
Anthropic shipped Claude Code v2.1.139 Monday May 11, the first feature-bearing release since v2.1.136 / .137 on Friday. The headline addition is the agent view (Research Preview): running claude agents returns a single list of every Claude Code session — live, blocked on you, or done — across all panes, worktrees, and Remote Control sessions. The release also adds /goal, which sets a completion condition that Claude keeps working toward across turns with a live elapsed-time / turns / tokens overlay, available in interactive, -p, and Remote Control. Other shipped features: /scroll-speed tuner with live preview, claude plugin details <name> for per-session token-cost projections, transcript-view navigation (? shortcuts, { / } to jump between user prompts, v toggle), hook args: string[] exec form (no shell, no quoting), continueOnBlock hook config for PostToolUse, MCP stdio servers now receive CLAUDE_PROJECT_DIR in their env, and subagent OTEL spans carry agent_id / parent_agent_id.
Per the changelog, the agent view and /goal are the first in-CLI primitives for the workflow Boris Cherny demoed at Code with Claude SF five days ago and that Altman publicly endorsed on Saturday (“kicked off a bunch of Codex tasks, came back at naptime to find them all completed”): queue parallel sessions, walk away, return later. /goal closes a real ergonomics gap — users had been wrapping the same behavior in shell loops or manual re-prompts. MCP stdio servers receiving CLAUDE_PROJECT_DIR is the smaller but more structural fix — it lets plugin authors scope external tooling to the current project without per-server hacks, the kind of plumbing that quietly unlocks plugin authors who were already at the edge.
Trial week 3 Day 1 (May 11) — Sutskever testifies live, says he spent ~1 year gathering board-requested evidence of Altman’s “consistent pattern of lying” including “undermining and pitting executives against one another,” his OpenAI stake worth ~$7B, confirms remaining board met with Anthropic post-ouster about a merger where Anthropic would take over OpenAI’s leadership, never a promise OpenAI would stay nonprofit; Nadella concludes earlier with “sort of amateur city” description of November 2023 board chaos and pre-investment text to Kevin Scott reading “I don’t want to be IBM and OpenAI to be Microsoft”; Bret Taylor takes stand third, describes “dire” ouster period and Zico Kolter as nonprofit-board AI safety committee chair; Altman testimony pushed to Tuesday May 12 morning
Today Updated 8:00 PM Bloomberg, CNBC, Courthouse News, GeekWire, MarketScreener, NBC News, ABC7, DNYUZ, Reuters
Week 3 of the Musk v. Altman trial opened in Judge Yvonne Gonzalez Rogers’s Oakland federal courtroom with three witnesses through Monday. Satya Nadella took the stand first: he told the jury Musk never contacted him with concerns about Microsoft’s $13B in OpenAI investments ($1B 2019 + $2B 2021 + $10B 2023), called the November 2023 board chaos around Altman’s ouster “sort of amateur city, as far as I’m concerned” with apparent “jealousies” among directors, and confirmed the cumulative $13B was never structured as a donation — the partnership had a commercial element from day one, with sharp compute discounts in exchange for commercial rights. A Nadella text to Microsoft CTO Kevin Scott from before the 2023 tranche surfaced: “I don’t want to be IBM and OpenAI to be Microsoft” — the 1980 DOS reference. Ilya Sutskever followed Nadella and delivered the trial’s biggest revelations to date: his OpenAI stake is worth approximately $7B (per Bloomberg, one of the largest individual shareholdings); he spent about a year preparing a document for the board gathering evidence of Altman’s “consistent pattern of lying,” including conduct described as “undermining and pitting executives against one another”; and he confirmed that after Altman’s November 2023 ouster, the remaining OpenAI board members met with Anthropic about a merger proposal in which Anthropic would take over OpenAI’s leadership. Sutskever also testified there was never a promise OpenAI would remain a nonprofit. Bret Taylor (OpenAI board chair, OpenAI Foundation chair) took the stand third, described the “dire” Altman-ouster period, walked the jury through OpenAI’s structure, and named Zico Kolter as chair of the AI safety committee (a nonprofit-board function, not for-profit). Altman’s testimony, originally expected later in the week, now begins Tuesday May 12 morning. Advisory verdict still expected week of May 18.
Per Bloomberg’s Sutskever stake reveal and the “consistent pattern of lying” framing reported by Reuters and MarketScreener, this is the single most damaging witness day Musk’s side has gotten: the OpenAI insider voice his lawyers most needed to mine for a corporate-mission-betrayal body blow just delivered it in court with a year of pre-collected receipts. The post-ouster Anthropic merger disclosure is the structural blow — it confirms the board considered handing OpenAI itself to its closest competitor, which damages the singular-AGI-mission framing OpenAI used to justify the for-profit transition. Nadella’s “amateur city” / “next IBM” lines are color; Sutskever’s testimony is the threat to the advisory verdict. Altman’s Tuesday-morning testimony just got considerably harder to script.
OpenAI grants the EU access to GPT-5.5-Cyber under its Trusted Access for Cyber (TAC) program — vetted European businesses, governments, cyber authorities, and the EU AI Office all get access to the model OpenAI restricted out of the box on April 30; Anthropic still holding out on Mythos for the EU after “four or five” Commission meetings, with the EU saying talks are “at a different stage” than the OpenAI agreement that’s now ready to ship
Today CNBC, Quartz, eWeek, WION, Benzinga, Winbuzzer
CNBC reported Monday May 11 that OpenAI agreed to grant the EU access to GPT-5.5-Cyber — the cybersecurity-tuned variant of GPT-5.5 originally launched on a heavily restricted basis to a small set of vetted partners after the April 30 Cyber model launch — via the company’s Trusted Access for Cyber (TAC) program. Authorized European businesses, governments, cyber authorities, and EU institutions including the EU AI Office can now request access for vulnerability identification, malware analysis, reverse engineering, and patch validation. Anthropic, by contrast, is still refusing to release Mythos to the EU despite EU finance officials publicly pressing the company. Per Quartz, the European Commission has held “four or five” meetings with Anthropic over Mythos access, but the talks are “not yet at the same stage as the solution we have on the table from OpenAI.”
Per CNBC, OpenAI just converted a regulatory headache into marketing posture — the EU is publicly pressing both labs for cyber-model access, OpenAI said yes within weeks, Anthropic is still in meetings without a deal. The structural irony: Anthropic’s withholding posture on Mythos was the moral high ground 14 days ago when Amodei told JPMorgan there’s a “6–12 month window for governments to patch before Chinese frontier models catch up.” Today it’s the talking point for OpenAI’s EU lobbying push and the next time Anthropic’s congressional testimony or DOD posture is up for debate, this delta gets cited.
Yann LeCun posts Monday on X pushing back on Silicon Valley monopoly framing — cites that Attention was invented in Montreal (Bengio lab, 2014), AlphaFold was built in London (DeepMind), and Llama was trained in Paris (FAIR Paris); argues the “US labs lead by months” line undersells the global character of frontier AI; comes the same week LeCun is positioning AMI Labs as an explicit non-Silicon-Valley alternative and the same week the US government (CAISI) struck new pre-deployment eval deals with Google DeepMind, Microsoft, and xAI
Today @ylecun on X, AI Pulse Daily
Yann LeCun posted Monday on X arguing that the “Silicon Valley labs lead the world by months” framing materially undersells where the science actually came from. The specific examples he cited: Attention was invented in Montreal (Yoshua Bengio’s lab, 2014); AlphaFold was developed at DeepMind London; Llama was trained at FAIR Paris under LeCun’s own former Meta team. The post followed his May 4 Axios sit-down where he positioned AMI Labs as a contrarian bet against the LLM-only Silicon Valley axis, and lands the same week the Trump administration (CAISI) struck pre-deployment eval arrangements with Google DeepMind, Microsoft, and xAI (May 5), the second US-only round of voluntary policy commitments. LeCun is not litigating chronology — he’s defending a non-US locus for frontier work as the AMI Labs go-to-market wedge.
Per LeCun, the geographic citations are the part to hold onto — Montreal / London / Paris, not Mountain View. His timing is deliberate: AMI Labs needs to be legibly outside the Silicon Valley axis to differentiate against the OpenAI / Anthropic / xAI / Google trio that just signed CAISI eval deals last week, and the “global origins” framing is the rhetorical foundation for that pitch. The line worth watching is whether European or Canadian capital (or even a sovereign anchor) starts citing this framing as cover for backing AMI Labs at a non-Silicon-Valley premium — that’s where the rhetoric becomes a fundraising thesis.
OpenAI launches Daybreak this afternoon — enterprise cybersecurity initiative built on GPT-5.5-Cyber + Codex Security with 21 named partners (Cloudflare, Cisco, CrowdStrike, Palo Alto Networks, Oracle, Zscaler, Akamai, Fortinet, Intel, Qualys, Rapid7, Tenable, Trail of Bits, SpecterOps, SentinelOne, Okta, Netskope, Snyk, Gen Digital, Semgrep, Socket); Codex Security builds codebase-specific threat models, inspects realistic attack paths in isolated environments, validates issues, proposes patches for human review; Greg Brockman announces on X; explicit Anthropic Project Glasswing / Claude Mythos competitor; broader industry / government deployment in “coming weeks”
Today OpenAI, Engadget, MacRumors, Decrypt, TestingCatalog, Yahoo Finance, Blockchain.news
OpenAI launched Daybreak this afternoon — its enterprise cybersecurity initiative built around frontier OpenAI models, Codex Security (the application-security agent already in the OpenAI dev suite), and GPT-5.5-Cyber under the company’s Trusted Access for Cyber program announced this morning for the EU. The product is aimed at developers, enterprise security teams, security researchers, and government-linked defenders for finding, validating, and patching software vulnerabilities earlier in the development cycle. Codex Security can build a codebase-specific threat model, inspect realistic attack paths, validate issues in isolated environments, and propose patches for human review. The launch includes 21 named partners spanning major security vendors: Cloudflare, Cisco, CrowdStrike, Palo Alto Networks, Oracle, Zscaler, Akamai, Fortinet, Intel, Qualys, Rapid7, Tenable, Trail of Bits, SpecterOps, SentinelOne, Okta, Netskope, Snyk, Gen Digital, Semgrep, and Socket. Organizations can request vulnerability scans now; broader deployment with industry and government partners is planned in the coming weeks. Greg Brockman announced the launch on X.
Per Engadget, Daybreak is the explicit OpenAI response to Anthropic’s Project Glasswing / Claude Mythos — same problem space, opposite go-to-market. The morning’s EU GPT-5.5-Cyber access story was the regulatory / lobbying maneuver; Daybreak is the commercial product launch built on the same model surface. Combined: OpenAI bracketed Anthropic’s morally-restrictive Mythos posture from both the policy side (granting the EU what Anthropic withheld) and the commercial side (commercializing the capability with 21 enterprise partners while Anthropic remains in 11-organization limited preview). Within 96 hours the strategic gap between the two labs on cyber has gone from “different timelines” to “different theories of the business.”