ALETH / AI-BRIEF / 2026-05-09 / COLOSSUS SWITCHES FROM GROK TO CLAUDEThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
Anthropic took all the compute at SpaceX’s Colossus 1, >220,000 Nvidia GPUs.
EU negotiators provisionally agreed to delay the AI Act’s high-risk rules.
Google DeepMind, Microsoft and xAI opened their models to US pre-release tests.
SAP agreed to buy Prior Labs, pledging >€1bn for a frontier AI lab.
Cerebras launched its IPO roadshow, seeking up to $3.5bn at about $40bn.
Browse by day:
Weekend: CAISI-DeepSeek
Monday: Anthropic firm · DeployCo · Cerebras · SAP · Musk trial · White House
Tuesday: US tests · Anthropic-Google · GPT-5.5 Instant · Nscale · DeepMind union
Wednesday: Anthropic-SpaceX · EU AI Act
Thursday: Akamai · Firefox · Roche · GPT-5.5-Cyber · NLAs
Friday: Teaching Claude why
Weekend (2-3 May)
US government testers found DeepSeek V4 Pro is the most capable Chinese model they have evaluated, about eight months behind the US frontier.
The Center for AI Standards and Innovation (CAISI) published the evaluation on 1 May. On its tests, which include non-public benchmarks, the model matched GPT-5, while DeepSeek’s own figures put it level with Opus 4.6 and GPT-5.4.
CAISI evaluation of DeepSeek V4 Pro · NIST
Monday 4 May
Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs launched an AI services firm to build Claude into the operations of mid-sized companies.
The standalone firm has Anthropic engineers embedded in its team, and General Atlantic, Leonard Green, Apollo, GIC and Sequoia Capital also back it. Anthropic chief financial officer Krishna Rao said enterprise demand for Claude “is significantly outpacing any single delivery model”.
Building a new enterprise AI services company · Anthropic
Anthropic Partners with Blackstone to Launch AI Firm · Blackstone
OpenAI raised >$4bn from 19 investors for The Deployment Company, a joint venture that will help businesses put its software to work, Bloomberg reported.
TPG, Brookfield, Advent and Bain Capital are among the backers. The Financial Times had reported a guaranteed 17.5% annual return for the private equity investors and super-voting shares for OpenAI, terms Reuters could not confirm.
OpenAI Finalizes $10B Venture With Private Equity Firms to Deploy AI · Bloomberg
Cerebras launched its IPO roadshow, marketing 28m shares at $115-125 each for a Nasdaq listing under the ticker CBRS.
The offering would raise up to $3.5bn at a valuation of about $40bn, Bloomberg reported. It is the chipmaker’s second attempt at an IPO. Revenue rose 76% to $510m in 2025, and Cerebras booked net income of $237.8m (a $75.7m non-GAAP loss).
Cerebras Systems Announces Launch of Initial Public Offering · Cerebras
Cerebras Systems Form S-1/A · SEC EDGAR
Nvidia Rival Cerebras Seeks to Raise $3.5 Billion in US IPO · Bloomberg
SAP agreed to buy Prior Labs and committed >€1bn over four years to turn it into a frontier AI lab for structured business data.
Freiburg-based Prior Labs builds tabular foundation models and will keep operating as an independent entity. SAP also agreed to buy Dremio, an open data lakehouse platform, in a deal expected to close in Q3. Terms of neither deal were disclosed.
OpenAI president Greg Brockman took the stand in Musk v Altman, where Musk’s lawyer used his diary to question a stake now worth about $30bn.
Brockman said no when asked whether he would give $29bn back to OpenAI’s nonprofit arm. A Sunday filing showed Musk texted Brockman about a settlement two days before the trial, warning that he and Sam Altman would be “the most hated men in America”. Judge Yvonne Gonzalez Rogers kept the text out of evidence.
On Wednesday jurors saw a video deposition in which former OpenAI chief technology officer Mira Murati said Altman falsely told her a new model could skip the deployment safety board. She said she sent it through the board anyway.
Application to introduce evidence of pretrial communication · CourtListener
Musk sought settlement with OpenAI two days before trial · CNBC
OpenAI president forced to read his personal diary entries to jury · Ars Technica
Mira Murati tells the court that she couldn’t trust Sam Altman’s words · The Verge
The White House is discussing an executive order for a working group on AI oversight, including pre-release model review, The New York Times reported.
The group would bring together tech executives and officials, and the White House briefed Anthropic, Google and OpenAI on the plans. A White House spokesperson told Politico that discussion of potential executive orders was “speculation”.
Report on White House plans to review AI models · The New York Times
White House mulls tighter controls on advanced AI · Politico
Tuesday 5 May
Google DeepMind, Microsoft and xAI agreed to let US government evaluators test their models before release.
The US Commerce Department’s Center for AI Standards and Innovation (CAISI) will run pre-deployment evaluations under the agreements. They build on CAISI’s earlier partnerships with Anthropic and OpenAI, which it has renegotiated. Microsoft signed a parallel agreement with the UK AI Security Institute (AISI).
CAISI Signs Agreements With Google DeepMind, Microsoft and xAI · NIST
Advancing AI evaluation with CAISI (US) and the AI Security Institute (UK) · Microsoft
Anthropic committed to spend $200bn on Google’s cloud and chips over five years, The Information reported.
That would be >40% of the $462bn Google Cloud backlog Alphabet reported at the end of March. Neither company commented.
Anthropic Commits to Spending $200 Billion on Google’s Cloud · The Information
Alphabet Form 10-Q for Q1 2026 · SEC EDGAR
OpenAI began rolling out GPT-5.5 Instant to all ChatGPT users as the default model, replacing GPT-5.3 Instant.
In OpenAI’s evaluations it made 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts in medicine, law and finance. OpenAI also introduced memory sources, which show what context shaped an answer and let users correct it.
GPT-5.5 Instant: smarter, clearer, and more personalized · OpenAI
Nscale will deliver >66,000 Nvidia Rubin GPUs to Microsoft at Start Campus’s site in Sines, Portugal, from late 2027.
The deal expands Nscale’s deployment of >12,600 Nvidia Blackwell Ultra GPUs for Microsoft in the campus’s first building. Nscale will invest €695m, including €465m in a second 200 MW building.
Nscale to Deliver 66,000+ NVIDIA Rubin GPUs to Microsoft in Portugal · Nscale
Google DeepMind’s UK staff asked the company to recognise two unions, partly over Google’s new deal with the US military.
The workers voted in April and, in a letter shared with The Guardian, asked Google to recognise the Communication Workers Union and Unite as joint representatives. Union officials said at least 1,000 workers would be covered. Google said that at this stage “there has been no vote to unionise”.
DeepMind workers in UK vote to unionize amid deal with US military · The Guardian
Wednesday 6 May
Anthropic signed a deal to use all the compute at SpaceX’s Colossus 1 data centre.
The deal gives Anthropic >300 MW of capacity, >220,000 Nvidia GPUs, within the month. Grok runs on a newer data centre, Colossus 2. The deal turns idle capacity into SpaceX revenue before its expected IPO. Anthropic doubled Claude Code’s five-hour rate limits for paid plans that day. SpaceX said Anthropic also expressed interest in developing multiple gigawatts of orbital compute together. Musk wrote it reserves “the right to reclaim the compute if their AI engages in actions that harm humanity”.
Higher usage limits and a SpaceX compute deal · Anthropic
New Compute Partnership with Anthropic · SpaceX
Elon Musk on providing compute to rival AI companies · Elon Musk
How Elon grew to love Anthropic · Axios
EU negotiators agreed to delay the AI Act’s high-risk rules to December 2027.
Council of the EU and European Parliament negotiators reached the provisional deal overnight. Rules for stand-alone high-risk systems would apply from 2 December 2027, and those for high-risk AI embedded in products from 2 August 2028.
The deal adds a ban on AI that generates non-consensual sexual content or child sexual abuse material, and exempts the machinery regulation from direct application of the Act. Politico called the industrial carve-out a win for Germany, whose Chancellor Friedrich Merz pushed for it to keep Siemens and Bosch competitive.
Council and Parliament agree to simplify and streamline rules · Council of the EU
EU clinches deal to roll back AI restrictions · Politico Europe
Thursday 7 May
Akamai said a US frontier model provider committed $1.8bn over seven years to its cloud, and Bloomberg reported the next day that it is Anthropic.
Akamai disclosed the commitment with its first-quarter results, in which cloud infrastructure revenue rose 40% to $95m. Chief executive Tom Leighton said it was “further validating our position as a key infrastructure provider in the AI economy”.
Akamai reports first quarter 2026 financial results · SEC EDGAR
Anthropic Inks $1.8 Billion Computing Deal With Akamai · Bloomberg
Mozilla said it fixed 423 Firefox security bugs in April, most found with Anthropic’s Claude Mythos Preview.
That compares with 31 a year earlier, TechCrunch reported. The Firefox 150 fixes included 271 Mythos Preview findings, 180 rated sec-high. The fixes include a 20-year-old bug in Firefox’s XSLT code.
Mozilla’s pipeline gives the model a part of the code, tells it there is a bug there and asks it to find the bug and build a test case.
Behind the Scenes Hardening Firefox with Claude Mythos Preview · Mozilla Hacks
How Anthropic’s Mythos has rewritten Firefox’s cybersecurity · TechCrunch
Roche agreed to buy US digital pathology company PathAI for $750m upfront and up to $300m in milestone payments.
The deal builds on a partnership the two began in 2021. PathAI will join Roche’s Diagnostics division once the deal closes, expected in H2 2026.
Roche enters into a definitive merger agreement to acquire PathAI · Roche
OpenAI began rolling out GPT-5.5-Cyber in limited preview to defenders responsible for securing critical infrastructure.
OpenAI built the variant for red teaming and penetration testing, where defenders need to confirm in a controlled environment that a flaw can be exploited.
Scaling Trusted Access for Cyber with GPT-5.5 and GPT-5.5-Cyber · OpenAI
Anthropic published natural language autoencoders, an interpretability method that translates a model’s internal activations into readable text.
In safety testing the method suggested that Claude Opus 4.6 and Claude Mythos Preview believed they were being tested more often than they said. It detected that awareness on evaluation transcripts where stated awareness was near zero, and on <1% of real Claude.ai transcripts from users who opted in to training. In an auditing game, an investigator using the method found a model’s implanted hidden motivation 12-15% of the time, against <3% without it.
Natural Language Autoencoders · Anthropic
Friday 8 May
Anthropic published research finding that teaching Claude why an action is right curbs misalignment better than training on examples alone.
Training on responses that turned down honeypot scenarios cut the misalignment rate only from 22% to 15%, while adding Claude’s reasoning about its values cut it to 3%. Every Claude model since Haiku 4.5 has scored perfectly on Anthropic’s agentic misalignment test, where Opus 4 resorted to blackmail up to 96% of the time.
Teaching Claude why · Anthropic
Aleth · Open Access Research