Daily Brief Oct 15, 2026
How we checked the items below
For readers who want to dig in. What happened, in brief, is on the Today tab.
This issue draws on the research brief we write ourselves, produced early on October 15, the deep dive published in the small hours the same day, and last night's sweep. The four main-line items date from October 13 to 15. Most items in the later sections date from October 13 to 14. The exception is the "From the archive" item. We swept 329 pieces overnight, and this issue uses 32 outside sources, each with a link to check.
1. [Evidence update] (deep-dive recheck, October 15) Our recheck of the CPU racks for running AI agents: four vendors announced or showed one at the same OCP summit, and the only public buyer, Meta, bought CPU capacity from AWS rather than a rack
Not proven yet： the share-follows-supply reading rests on one Intel filing alone
How we checked
We read the AMD and NVIDIA originals. The Intel, Arm, Meta and DeepSeek parts come from the deep dive's compilation of vendor press releases, Intel's 10-Q and the arXiv paper. Those original documents have so far been compiled by the deep dive and not yet rechecked one by one. Sandbox utilization in live AI agent products has no public figure.
2. [Today] (event date October 13) Venture investor Tomasz Tunguz works out that five frontier models within five intelligence points of each other differ 3.8x in cost per task, with Gemini 4 Argon the cheapest
Not proven yet： the cost per task is Tunguz's own calculation; the evaluator's page doesn't list it
How we checked
We read Tunguz's post in the original and checked the quotes word for word. Artificial Analysis's public model pages list only intelligence scores, not the same set of cost-per-task figures. The 44% is his own estimate. Signal65's figures come from an archived copy of the post, and its methodology page hasn't been reviewed. Signal65 is the measurement unit of the analyst firm Moor Insights & Strategy. Both readings are a single organization's own calculation.
3. [Today] (event date October 13) A Delaware jury finds Arm breached its license agreement with Qualcomm and interfered with Qualcomm's Meta business, awarding only US$2; whether Qualcomm can stop paying royalties is still for the judge
Not proven yet： the judge hasn't ruled on whether Qualcomm can stop paying royalties for five years
How we checked
We read the Reuters story and the Qualcomm and Arm statements in full. All three agree on the one breach claim that succeeded and differ on how they count the claims won. The Reuters story was read as relayed on a radio station's site, and the version on Reuters' own site hasn't been compared. The original gives the royalty amount only as "billions". Neither the judge's ruling nor the antitrust decision has a date yet.
4. [Today] (event date October 14) OpenAI customer story: the quant startup Third Moment self-reports running 10,000 agents at once for 36 to 48 hours each, with an outer program deciding whether to retry, terminate or escalate to a human
Not proven yet： the 10,000 figure is the customer's own account, relayed by its vendor
How we checked
We read OpenAI's case page in the original and checked the quotes word for word. It is a vendor marketing page. Every figure is the customer's own account, with no third party and no results figures. The page doesn't say whether the 10,000 agents divide work among themselves; from the description they look like each running its own investigation, which is our reading.
[Evidence update] (filed October 14) ASML's 6-K filing with the US Securities and Exchange Commission says its currently planned capacity will not fully meet customer demand in either 2027 or 2028. ASML also says it is studying adding capacity beyond the 30% expansion it announced earlier.
How we checked
[Today] (October 14) Marvell starts sampling two 224G-per-lane scale-up copper connectivity chips: a 3.2T retimer and a second-generation active-copper-cable DSP. The DSP is built on 3 nm and draws 12.8 W.
How we checked
[Today] (October 14) A Databricks engineering post says a Qwen 3.8 27B fine-tuning run went from 7 hours 44 minutes to 1 hour 17 minutes, about 6x faster, even though GPU utilization already read 98%.
How we checked
[Today] (post October 14) Compute-governance researcher Lennart Heim says a 20x compute advantage buys only about six months of lead in model capability. If distillation is blocked as well, the lead is about 12 months.
How we checked
Both numbers are his estimates, and the posts attach no underlying analysis.
[Today] (clip October 14) Ai2 researcher Nathan Lambert gives the example that an average OpenAI researcher can draw on about 5,000 H100-equivalents of compute at any time.
How we checked
The clip doesn't give the venue or date of the conversation.
[Today] (October 14) METR, a third-party evaluator, proposes an external assessment scheme for loss-of-control risk in frontier models, and says the scheme itself gives no assurance of safety.
How we checked
It estimates that two months at the same scope would cost more than US$10M, which is its own estimate.
(METR, 2026-10-14) (Beth Barnes, 2026-10-14) (METR, 2026-10-14)
[Today] (October 14) Vercel publishes the results of its US$1M sandbox bounty: a white hat found a Linux kernel zero-day that broke through its virtualization isolation, and Vercel CEO Guillermo Rauch says the flaw is patched.
How we checked
All figures are Vercel's own, and whether the same vulnerability breaks through other providers has no third-party check.
(Vercel, 2026-10-14) (Guillermo Rauch, 2026-10-14) (Vercel blog, 2026-10-14)
[Today] (October 14) Kiro, AWS's development environment for coding agents, adds a local sandbox: even after a command is approved to run, the operating system still limits which files it can read and write.
How we checked
The agent program itself still sits outside the sandbox, and the effect has no third-party measurement.
Sources & accounting
The past 24 hours. 329 new pieces came in from October 14 to 15: 135 company or personal blog posts, 124 social-platform posts, 50 papers from other sources, 9 subscription newsletters, 6 arXiv academic papers, 2 show transcripts, 2 macroeconomic data releases and 1 company filing. Of those 329, we read 107 and screened out 84 as off-topic. The other 137 haven't been read yet, and 1 failed to fetch. Of the 137 unread, 66 are blog posts, 50 are papers from other sources and 21 are social-platform posts; those 50 papers arrived in a single delivery from one paper-digest subscription, and none has been read. arXiv's servers were overloaded last night, so its papers couldn't be fetched, and the academic section doesn't run this issue. Four more sources are logged as unfetchable: Data Center Dynamics, the Meta AI blog, and the investor news pages of Broadcom and Qualcomm; any announcements from them that evening fall outside this issue's sources. The Arm, Qualcomm, Reuters, Tunguz and OpenAI originals behind the main line, and the ASML filing in the chips section, are originals we read directly today.
One-time catch-up. No new one-time catch-up of older material today.
A note on source concentration. ⚠️ In main-line item 1, the Intel, Arm, Meta and DeepSeek readings come from our deep dive's compilation, and the original documents haven't yet been rechecked one by one. ⚠️ Both readings in main-line item 2 are a single organization's own calculation. ⚠️ The stories carrying "Also happened" items 1, 2 and 6 are paywalled in full. ⚠️ Both Product moves items rest on the vendor's own account.
What is missing today. What would most affect judgment is sandbox CPU utilization in live agent products: no company has published one. Beyond that, Artificial Analysis's underlying cost-per-task table, Signal65's methodology page, the full Bloomberg and Financial Times stories on Anthropic, and the verdict story on Reuters' own site haven't been reviewed; Marvell's press release gives no customer list.
The sources we track. Our long-term roster has 529 named sources: 302 on social platforms, 90 shows, 51 news outlets, 48 blogs, 48 paper authors and 46 newsletters, with the rest spread across earnings reports, keynotes and other channels. ⚠️ Those are counts of tracked sources, a different population from the 329 new pieces above; they can't be added or compared. Representative names: on social platforms, Elon Musk, Alexandr Wang, Ethan Mollick, Ben Bajarin and Paul Triolo; among newsletters, Zvi Mowshowitz, Ben Thompson, Timothy B. Lee and Eric Newcomer; among shows, Sequoia's Training Data, Latent Space, The Cognitive Revolution and Taiwan's Gooaye. This issue uses 32 outside sources in the body. We count only body links not on our own domain. We added 7 Taiwanese sources in September, bringing the Taiwanese total to 127; Taiwanese media, company official channels in Taiwan and local reporters there count; foreign outlets relaying Taiwanese media do not.