Four threads run through today’s AI news, and the first is who gets the chips. Sundar Pichai says Google’s cloud arm gives up some revenue so DeepMind gets compute first, while Google and Meta built TorchTPU to run PyTorch on Google’s TPUs.
The second is the price of an answer. DeepSeek says one cluster serves 155 trillion tokens a day, and Mistral is selling a full phone-call agent from 4 cents a minute.
The third is agents on the job. Ramp built its own platform in six weeks, Databricks says memory lifted an agent’s pass rate from 22.3 to 68.4 percent, and Microsoft says Storm-0940 pointed ARTEX at more than 2,250 tenants.
The fourth is checking the work. METR calls Anthropic’s risk low but credits weak models, and 637 surveyed researchers say testing ideas is now the slow part.
Who Gets the Chips: The Scramble for Compute
Google’s CEO admits his own cloud customers wait behind DeepMind, and a dinner of compute buyers says money and power set the limit. Meanwhile, two rivals work on making the chips easier to switch between.
- Pichai admits Google Cloud is giving up sales so DeepMind gets chips first. Sundar Pichai told Big Technology that Google’s cloud business forgoes short-term revenue because frontier research gets compute first. He said Gemini 4 Argon handled almost 25 trillion tokens in its first 72 hours and that Google is still capacity constrained.
- Six people who buy GPUs for AI labs say money and power are the limit. Nick Rubin, who co-hosted the dinner, writes that almost everyone there could sell all the capacity they bring online, so financing and electricity decide who grows. He doesn’t say who attended, so the account rests on his word.
- Google and Meta let PyTorch code run on TPUs by swapping a device name. TorchTPU is open source, and the team says a basic port can mean changing one line, though peak speed still takes tuning. The benchmarks are Google’s own and none compares against Nvidia hardware.
- OpenAI will sell its models through the cloud of its biggest rival. OpenAI’s models and Codex are headed for Google Cloud’s Model Garden, and customers can pay with money already committed to Google. The announcement names no models and no prices, and says Microsoft remains OpenAI’s primary cloud partner.
The Cost of an Answer: Serving AI at Scale
DeepSeek published two looks at how it serves its model, one claiming big money per chip and one handing over an incomplete engine. The voice-agent makers arrive with price tags attached.
- DeepSeek says one cluster handles 155 trillion tokens a day at $38 per GPU-hour. The figures come from DeepSeek’s own Zhihu post, which credits split hardware, a three-day cache for agent traffic, and faster guessing. Nobody outside the company has checked them.
- DeepSeek open-sources its serving engine but leaves out the cache and load balancer. DeepInfer covers offline inference for DeepSeek-V4.1-Flash on Nvidia and Huawei chips. A team still has to build the memory cache, the split between reading and writing machines, and the traffic balancing itself.
- Mistral prices a full phone-call voice agent at 4 cents a minute. Voxtral 3 bundles transcription, a voice model, and an agent kit, with the cheapest setup at $0.04 a minute. The benchmark claims are Mistral’s own, and the announcement doesn’t say whether the weights can be downloaded.
- ElevenLabs offers voice agent models that drop into code written for OpenAI. Eleven Realtime comes in a Mini and a Max version, and a team can switch by changing the endpoint and the model name. The product page publishes no pricing, response times, or independent tests.
Agents on the Job: Memory, Crashes and an Attack
Companies are building the machinery that keeps agents working, and Microsoft says a hacking group put one to work on offense.
- Ramp built its own agent platform in six weeks, with AI writing the first version. Engineer Sam Kronick says experienced engineers reviewed every change, and Ramp still needs people on call to keep Laminar running. The post gives no cost figures.
- Vercel’s durable-harness lets a crashed coding agent resume where it stopped. The open-source package logs each model and tool call, so finished calls aren’t billed twice, and risky tools such as a refund aren’t rerun after a crash. It works with Pi, OpenCode, Codex, and Claude Code.
- Databricks says agent memory lifted pass rates from 22.3 to 68.4 percent. The test set is small and Databricks built it from its own Genie conversations. The best rival it tried, Mem0, reached 43.9 percent.
- Microsoft says China-linked hackers let an AI tool run a supply-chain attack. Storm-0940 pointed the open-source ARTEX tool at more than 2,250 tenants, and some intrusions reached admin access in under three hours. Microsoft says most of the targeting failed and doesn’t say how many tenants were breached.
Checking the Work: Where AI Still Waits on Proof
Safety reviewers, a Google-backed survey, and a former satellite engineer land on the same snag: producing answers has outrun confirming them.
- METR agrees Anthropic’s risk is low, but credits weak models, not safeguards. METR says the August report’s monitoring tests are a poor stand-in for the real system. In one incident the monitor flagged about 1 percent of misaligned actions until the model’s reasoning was removed, and then roughly half.
- OpenAI says one score for grading AI watchdogs can rank them wrongly. The alignment team adds a pair-matched comparison and a retry-based metric after finding its g-mean² can favor the wrong model. A low risk number can also just mean the system rarely answers.
- Testing ideas, not having them, is now the slow part of AI-assisted science. In a survey of 637 researchers, AI saves about seven hours a week, and 88 percent spend more than 10 percent of that time checking its output. Annelies Gamble’s write-up adds that nearly half report a growing backlog of untested hypotheses.
- A former satellite engineer says AI will design hardware long before it builds any. Christian Keil argues that real objects vary, test time can’t be compressed, and process knowledge is scarce. His case is a hypothesis, with no measure of how fast those barriers are falling.
- A paper proposes AI that keeps verified facts apart from its reasoning. Metis Reasoning borrows from AlphaGo’s design, but the whitepaper reports no experiments, benchmarks, or working system.
Today’s Quick Hits
- Free software spreads one AI model across many servers. llm-d, an open-source project for running models on Kubernetes clusters, offers tested deployment recipes that work with vLLM and SGLang across GPUs, TPUs, and CPUs. Its 3x throughput claim is the project’s own, with no baseline given.
- OpenAI buys Orchestra Bio to run drug programs with AI. OpenAI bought Orchestra Bio, whose agents track a drug program’s studies and timelines, and will fold it into Rosalind, its life-sciences line. Financial terms weren’t disclosed, and the announcement offers goals rather than customer results.
- ChatGPT adds a PIN lock for private conversations. ChatGPT’s new Locked Chats put conversations behind a four-digit PIN or Face ID, in a separate section with its own memory. OpenAI says it isn’t end-to-end encryption, and resetting the PIN deletes what it protected.
- SparkLLM ships a small open model for phones and laptops. SparkLLM released Spark-X2.6-4B under Apache 2.0, a four-billion-parameter model meant for local hardware. The model card says it matches some 9B models, but states no scores and cites no independent test.
- Meta’s Muse hit 10.2 million US downloads in 40 days. Meta’s Muse reached 10.2 million US downloads in 40 days, against 3.7 million for ChatGPT, per Sensor Tower. ChatGPT still leads worldwide, and Muse is offered only in the US and Canada.
- A Berkeley essay says Coxon’s warning left the safety crowd unmoved. Rufus Knuppel’s essay says Jacob Coxon’s resignation from Anthropic reached big podcasts and politicians but left the safety community barely changed. The evidence is conversation and one unnamed researcher, so it reads as opinion.