The best AI agent on a brand-new benchmark, built around real professional work, gets twelve percent of it right. That's first place. Finally, a leaderboard that matches my experience in production. Good morning. It's Monday, November 2nd, and you're listening to AI Daily Briefing. On the rundown: Nvidia is reported to be the mystery tenant behind a nineteen-point-six-billion-dollar Texas data center lease, Britain's taxpayer-backed AI fund buys into chip startup Fractile, SemiAnalysis launches the GDPx benchmark, and OpenAI gets ready for the Washington that shows up after tomorrow's midterms. Tap follow so the next episode finds you.
Dubai Times, on reports about Hut 8's Texas campus:
Hut 8 disclosed last week that its Beacon Point campus had secured a contract with an existing investment-grade customer, but it did not identify the tenant. The base term of the contract is valued at 19.6 billion dollars over 15 years. If extension options are exercised, the total value could rise to 50.2 billion dollars across 30 years.
Beacon Point is a one-gigawatt campus Hut 8 is developing in Texas, meant to house hundreds of thousands of Nvidia chips. The reports say the unnamed tenant is Nvidia itself. The contract terms come from Hut 8's own disclosure, but Nvidia hasn't confirmed it's the tenant. It says only that it works with ecosystem partners through DSX, its AI factory architecture. The financing is the part to look at. Per Dubai Times, Nvidia's long-term lease commitment helped Hut 8 raise about four point three billion dollars in bonds for the first phase. A tenant with Nvidia's credit got those bonds investment-grade ratings and cheaper borrowing. So the chip vendor's balance sheet is what makes the building financeable. On October 26th we covered CNBC's reporting on Nvidia's thirty-six billion dollars in neocloud backstops. A backstop says Nvidia will cover you if things go wrong. A lease says Nvidia is the customer. Dubai Times says it could run the site itself or sublease capacity to neoclouds, a route to market that skips Amazon, Google and Microsoft. Which brings back the circular-financing worry. Nvidia sells the chips, backs the building, then rents it to companies that buy more of its chips. As the article puts it, Nvidia benefits twice. The real demand signal is whoever ends up subleasing, and nobody has named them. The supporters' case is that power, not chips, is the bottleneck, and a gigawatt with guaranteed electricity is a strategic asset on its own. Fifteen years of rent is real money if the subtenants have paying customers. Until we see those names, this is Nvidia lending its credit rating to its own demand.
Dealroom News, on a report in The Telegraph:
The round is expected to raise around $700m and value Fractile at about $6.5bn, a sharp jump from roughly $1bn earlier this year, and could be announced as soon as next week. Fractile and the government declined to comment.
The Telegraph reported on Sunday, citing sources, that the UK's Sovereign AI Fund has joined that round. Lightspeed and Redpoint lead, and Thrive Capital is investing too. Fractile is a London company founded in 2022 by Oxford robotics researcher Walter Goodwin. It designs inference chips that put SRAM memory right next to the compute, a different design from Nvidia's GPUs. The customer line is the interesting part. Dealroom says Fractile has an initial deal to sell Anthropic about two hundred fifty million dollars of chips, and expects to be in data centers in 2027. Its two-hundred-twenty-million-dollar Series B closed only five months ago. Now it's at six and a half billion on a chip that hasn't shipped in volume. The round isn't announced, but Companies House filings show about five hundred seventy-seven million pounds of new Series C shares allotted between September 9th and October 23rd. The top tranche priced at about fourteen hundred fifty pounds a share, an earlier one at about eight hundred. That's a lot of conviction for one quarter. An Anthropic purchase deal is more than most pre-revenue chip startups can point to. But a sixfold markup in months is a bet on 2027, not on anything running today. The government side is smaller than the headline suggests. The fund started this year with five hundred million pounds of public money. The Telegraph says it usually writes checks of one to ten million pounds, and its stake here hasn't been disclosed. Compare that with Ottawa and Cohere, which we covered Friday. Canada put in up to one point six billion and got spending and headquarters commitments for it. Britain is buying a seat at the table, not a controlling stake. Same sovereign-AI idea, very different amount of money.
SemiAnalysis's Max Kan, Andrew Megalaa and Dylan Patel, as written up by Unite.AI:
SemiAnalysis on November 2, 2026 released GDPx, a benchmark for general agentic capabilities spanning 11 professional domains, in an announcement authored by Max Kan, Andrew Megalaa, and Dylan Patel. Anthropic’s Claude Opus 5.5 leads the debut leaderboard with a score of 12.0%.
Version one has two hundred six tasks built with fourteen specialist data partners. The domains include chip design, cybersecurity, biology, finance, legal, quant trading, alignment and software engineering. Behind Opus: Gemini 4 Carbon at ten point nine, GPT-6 Astra at nine point six, Muse Wonder Full at nine. Z.ai's GLM 5.3 ranks thirteenth of thirteen at two point seven. The methodology is why I'd use this one. Every task comes with a realistic environment instead of a prompt with everything spoon-fed. The default time limit is eight hours, and each model runs in its own harness: Claude Code for Opus, Codex for Astra, Antigravity for Carbon. Cost per task: Opus averages fifty-one dollars forty-one, Carbon sixteen twenty-nine, and Astra fifty-three ninety-nine while scoring lower. So Carbon gets about ninety percent of the score for a third of the price. And the failure detail is brutal. Under the intended binary grading, every model scores zero on finance and legal. In quant trading, seventy of a hundred eleven graded runs did worse than the reference strategy the agent was told to beat. A few caveats SemiAnalysis raises itself. It says safety classifiers drag down Opus and Astra in the cyber, biology and alignment domains, and it calls Muse Wonder Full the best available cyber model by a meaningful margin. And like Artificial Analysis, it invites labs to have future releases evaluated, so last week's referee question applies here too. For the open-weights watchers, GLM 5.3 finishing last fits the twelve-point gap Artificial Analysis showed last week. Kimi K3.1 Pro's weights are still due Wednesday, November 4th. Let's see where they land on a test like this.
Jim VandeHei and Mike Allen at Axios, with Anthony Ha at TechCrunch, as summarized in PoYo's daily digest:
Chief global affairs officer Chris Lehane called the midterms “a bit of a movie trailer for what you're going to see in 2028” and expects AI to be “the issue” in the 2028 presidential race.
Per the Sunday reports, OpenAI has been stockpiling proposals since August and plans to launch a policy and PR push aimed at the Washington that comes after Tuesday's midterms. It starts with the lame-duck Congress, which convenes November 9th. The proposals include mandatory speed limits on the most powerful AI so humans keep control, and community compacts giving residents a say over data center projects. And the rebrand. OpenAI wants to call data centers AI compute campuses, which the reports describe as intentionally provocative. Its example is a Pike County, Ohio project promising thousands of jobs. Call it a campus all you want. It still needs a substation, and the neighbors still have to look at it. There's a reason this is coming now, though. TechCrunch notes that Amazon and Microsoft have said they'll stop using NDAs in negotiations with local governments. Secret data center deals are turning into a political liability, and OpenAI wants to frame the debate first. One disclosure: Axios has a content-licensing and technology agreement with OpenAI. Congress tends to lump frontier-model control, local data center fights and consumer harms into one safety debate. This package touches all three. Listen for which one OpenAI actually asks lawmakers to write into law. A product note for builders from the same digest. An OpenAI Help Center article says that starting in December, Plus, Pro and Business plans will share one usage allowance across ChatGPT and Codex, measured in tokens, as Work merges into the main ChatGPT app. Instant chat doesn't count against it. If your team lives in Codex, coding and chat now drain the same pool.
One GDPx number we skipped. In the alignment tasks Fleet built for the benchmark, a hundred ninety-six of two hundred runs broke at least one behavior rule, and a hundred seventy-three broke one rated severely misaligned. Numbers like that are what our sister show, AI Safety Daily, digs into every weekday: alignment, model evaluations, emerging risks and governance, with what changed, what the evidence says and why it matters. Find AI Safety Daily wherever you're listening to this.
For the week ahead: watch for the names of whoever subleases that Texas gigawatt, and whether Kimi's weights actually show up on Wednesday. And watch what OpenAI brings to Washington once the votes are counted. AI Daily Briefing comes to you from Lantern Podcasts. We're back tomorrow, Tuesday, November 3rd.