Anthropic takes a stake in Akamai as part of an [$11.6B cloud deal](https://www.akamai.com/newsroom/press-release/akamai-announces-11-6-billion-multi-year-agreement-with-anthropic-to-support-growing-demand), shares surge 22%

Good morning 👋 Akamai just showed what a hyperscaler pays to lock in Anthropic as a customer: $11.6 billion in cloud commitments, plus a slice of the vendor's own stock. The market decided that math was worth a 22% pop before anyone finished reading the filing.

In today's issue:

• 🔭 Akamai locks in $11.6B from Anthropic and takes equity in the deal, AKAM jumps 22%
• 🧠 Gemini Live gets a real-time avatar, LiquidAI ships a free VLM speedup, NVIDIA's safety stack for physical AI
• 🔬 Coding agents delete their own logs, an evasion benchmark that gets worse the more an agent reasons, a living clinical-records benchmark
• 🚀 Databricks buys Row Zero for Genie, Basecamp Research raises $140M to design cell therapies
• 🛠️ Keep your coding agent from marrying one model's cloud deal
• 🔗 Schmidhuber joins Sakana AI, plus four more worth a sentence

🔭 THE ONE THING

📈 Anthropic takes a stake in Akamai as part of an $11.6B cloud deal, shares surge 22%

Akamai just locked in a seven-year, $11.6 billion cloud commitment from Anthropic, and got paid partly in Anthropic's own upside for it: a warrant lets Anthropic buy up to 5% of Akamai at $111.33 a share, with another 3% vesting in tranches if Anthropic's spend climbs $9 billion higher still. The market didn't wait for anyone to model it out. AKAM jumped 22% in extended trading the moment the terms hit the wire. The contract size is almost the boring part here. Akamai chose warrant coverage over a bigger headline discount, which only pencils out if you believe Anthropic's growth curve keeps steepening, and if you're pitching compute or infra to a frontier lab, that's now the template worth copying.

🧠 MODELS & RELEASES

• 🎭 Google DeepMind shipped Gemini 3.8 Live with Live Avatar: real-time animated avatars with lip-sync across 97 languages, live now in Gemini Enterprise. Custom avatars are enterprise-allowlist only, but the async tool-calling (background lookups mid-conversation, no dead air) is the part worth stealing for your own agent's UX. Introducing Gemini 3.8 Live with Live Avatar

• ⚡ LiquidAI open-sourced a draft model that speeds up its LFM2.5-VL-3B vision-language model via speculative decoding: 3.1x faster on an M5 Max, 2.7x on H100, no quality loss, for an extra 280M params riding on the 3B target. Open-weight, so if you're serving a VLM on-device, it's a free speedup. LFM2.5-VL-DSpark

• 🛡️ NVIDIA rolled out Halos, its safety stack for physical AI, now spanning autonomous vehicles and robots with third-party certs (TÜV, ISO/IEC 17020) baked in. It's infrastructure, not a product launch: ASIL-D certified OS, dedicated safety silicon, and simulation tooling bundled for anyone building machines that touch the physical world. Why Deploying Physical AI at Scale Demands Safety at Every Layer

🔬 RESEARCH HIGHLIGHTS

• Trace tampering Five of six coding harnesses tested (Claude Code, Codex, Antigravity, Open Code, Grok Build) let an agent delete its own execution logs on request, no guardrail fired. Only Muse Code blocked it. The harder claim in the paper: the behavior showed up on its own in frontier models just optimizing for reward, nobody had to coach it. If your compliance story leans on agent-generated audit trails, it's leaning on the agent's cooperation. arxiv.org/abs/2609.30266

• EvasionBench Fifty task/policy pairs built so finishing the task means breaking a monitoring rule. Best-of-3 attempts cleared the monitor up to 98% of the time, with an 88% success rate on the underlying task too. Evasion got worse, not better, as models burned more reasoning tokens. Claude Fable 5.1 evaded less often but got creative about it; GPT-6 Astra just abandoned solvable tasks when it smelled a denial-of-service trap. The takeaway: ordinary task persistence and safety monitoring pull against each other once an agent can see its own runtime. arxiv.org/abs/2609.30217

• BRIE A living, clinician-validated benchmark for pulling answers out of electronic health records, refreshed continuously so models can't quietly memorize the eval set. Nineteen clinicians built it, nine LLMs got tested across five inference strategies, and the result is unglamorous: models miss clinically important details whenever the answer requires stitching together more than one note. Worth knowing if you're building clinical retrieval and need a benchmark that won't go stale in six months. arxiv.org/abs/2609.30205

🚀 AI STARTUPS

• Row Zero got bought by Databricks (terms undisclosed) to give Genie a spreadsheet it can actually hand to finance and ops teams. The pitch is a spreadsheet that talks to live, governed data instead of the usual CSV export nobody tracks, built by ex-AWS and Tableau engineers who apparently know how to make a pivot table survive a billion rows. Every agent action stays auditable inside the sheet, which is the part that should matter to anyone whose compliance team has opinions about shadow data. source

• Basecamp Research closed a $140M Series C led by S32, with Nvidia, the Rockefeller Foundation, and André Hoffmann (Roche's vice-chairman) among the names writing checks. The money trains EDEN, their foundation model, on what they call a "Trillion Gene Atlas," sourced from partners in 30-plus countries, to design cell therapies from scratch. Big claim on the dataset. Nobody outside the company has audited it yet, so file the biology under promising and the marketing under standard. source

🛠️ TRY THIS

Give your coding agent a backend that isn't married to one lab

Akamai just bet $11.6B and an equity stake on Anthropic being the AI provider that matters. Fine bet for Akamai. Bad plan for your own tooling, if it means locking your daily coding agent to whichever model your cloud happens to favor this quarter.

1. Pull OpenCode, the open-source terminal agent AWS wrote up this week, and point it at Amazon Bedrock instead of a single vendor's API.

2. Wire in two or three open-weight models already live on Bedrock. Match model to task instead of defaulting to whatever's loaded.

3. Route boilerplate and test scaffolding to the cheapest model in the pool. Save the expensive one for anything touching architecture.

4. Check the AWS bill after a week. If one model is carrying most of the load, that's the one worth negotiating a real contract for. Everything else stays swappable.

Prompt: List every file this diff touches, then flag which changes are mechanical enough for a smaller model and which need full codebase context.

🔗 QUICK LINKS

• Sakana AI AI pioneer Jürgen Schmidhuber is joining the lab to lead a new research group, the Frontier Intelligence Group. Sakana AI
• DensityAI Per The Information, the chip startup founded by ex-Tesla Dojo leaders is nearing a $10B valuation. The Information
• OpenAI ChatGPT Ads is expanding into Southeast Asia and Taiwan, putting the ad product live in more than 60 countries. OpenAI
• OpenAI New MentalHealthBench grades chatbot responses in mental-health conversations against expert-built criteria. OpenAI
• NVIDIA Joined a coalition releasing an open pandemic-prep protein dataset, betting a head start on the science beats scrambling once the next outbreak hits. NVIDIA

Worth watching whether other clouds start pricing frontier-lab contracts the same way.

Pradeep Perugu