On October 1st, Google's new frontier model was look-but-don't-touch. Now it has a price sheet. And that price sheet has an expiration date. This is AI Daily Briefing, and today is mostly silicon and concrete. Microsoft wants its own chips serving Copilot, Anthropic is reportedly close to a site in Osaka, and mathematicians are still working through a proof that OpenAI's agents wrote. If today's show was useful, follow us wherever you're listening — the next one will be waiting.
Ryan Whitwam, writing in Ars Technica:
For developers accessing Google’s models via the paid API, the company will happily sell you some Argon tokens. Pricing is temporarily set at $2 per million input tokens and $10 per million output tokens, with a 95 percent discount for cached tokens. This deal is only good through November 30, when the rate will jump to $4 per 1M input tokens and $20 per 1M output tokens.
We covered the Gemini 4 Argon announcement on October 1st, back when nobody outside Google could use it. As of Tuesday it's on the API, and the model card went up the same day. Two dollars in, ten out. Until November 30th. Then it doubles. And four in, twenty out is the same rate Meta put on plain Muse Wonder, which we covered yesterday. So frontier list prices are lining up, and Google's using a seven-week discount to get you wired in first. That's what I'd plan around. Nobody rebuilds an agent pipeline in seven weeks. You build on the intro rate, and in December the bill doubles. What I'd actually test is the ninety-five percent cache discount, for agents that keep re-reading the same repo. On the consumer side it's gated hard. The Gemini app needs AI Pro or Ultra. Antigravity, the coding product, is Ultra only, and Google wouldn't tell Ars when or if that changes. Google also says Argon ranks first on the Vals Index. That's Google's own claim, and I'd wait for outside evaluators. One Hacker News commenter said the place Argon can really be tested sits behind a bigger paywall than the one they already pay for. It's the Wonder Full complaint all over again. The strongest coding setup is the hardest one to get your hands on.
Kyle Baird, writing in The Latent:
The company plans to expand the existing Maia chip line to "multi-gigawatt deployments," according to a Monday release. It did not give an exact target for capacity or when the rollout could be completed.
On Monday, Satya Nadella called it Microsoft's biggest bet yet on its own silicon. Microsoft says that starting in 2027, Maia becomes the primary inference system for its products. But there's no gigawatt figure and no completion date. The hardware is real, though, and there's a technical paper. Maia 300 has eight HBM4 stacks, eighteen terabytes a second of bandwidth, and a 950-watt TDP. A rack holds ninety-six accelerators at 167 kilowatts. And Maia 200 is already serving Copilot, running OpenAI models, MAI models and open-source families. That's production traffic. It's not a slide deck. The headline number is forty percent more tokens per megawatt than leading GPU-based systems. That's Microsoft grading its own chip against a competitor it doesn't name. Nvidia shares still fell about 1.7 percent Monday morning. I'll give them this: tokens per megawatt is the right fight to pick. If that forty percent holds on long-context, decode-heavy agent work, Copilot's serving costs look very different. Notice it's an inference play. Microsoft is still an Nvidia customer. Baird points to the July Mistral partnership built on Vera Rubin GPUs. What Microsoft wants to own is the serving side, and Nvidia isn't going anywhere yet.
John Marshall at WebProNews, following a scoop from Nikkei Asia:
NTT Data Group stands on the verge of a landmark agreement. The Japanese technology giant is in final negotiations to construct Anthropic’s first dedicated data center in the country. Located in Osaka, the facility could carry a construction price tag of up to 1 trillion yen, or $6.3 billion. Partial operations might begin as early as 2028.
Up to five hundred megawatts in Osaka. The grid can't deliver that quickly, so the plan starts with on-site generation and fuel cells. Bring your own power plant. That's how this works now. Let's be clear on where this stands. It's final negotiations, reported Tuesday by Nikkei and then by Reuters. Neither NTT nor Anthropic commented. The goal is to sign by the end of Japan's fiscal year, March 31st, 2027. And six point three billion only covers construction. Add the GPUs and the piece says the total reaches several trillion yen. This goes in the ledger we've been keeping. Anthropic's IPO prospectus showed at least 518 billion dollars in compute commitments over a decade, and Data Center Dynamics counts at least 14.8 gigawatts of capacity signed since last October. Most of that is in the US. Last month it was a first lease in Australia. Now Japan. For once I don't have to ask who the tenant is. It's Anthropic. And NTT needs that anchor. It's trying to grow its Japan capacity from about 300 megawatts to a gigawatt by fiscal 2033, and this one site could cover a big chunk of that.
Joseph Howlett, writing in Scientific American:
What’s important, and perhaps a harbinger of math’s future, is how long it took experts to realize all this. When OpenAI revealed its landmark achievement on September 8, company officials clearly and repeatedly stated that their proof followed the Spanish program. Their combative exchange with Buckmaster before the release seems to suggest that they believed as much. And they weren’t initially certain whether his work had been incorporated in their model’s training data.
Some background. On September 8th, OpenAI posted a 166-page Navier-Stokes proof built by a swarm of about ten thousand copies of an internal model. That's roughly eighty-eight hours and a hundred thirty billion tokens. And OpenAI said the proof followed an approach from two Spanish mathematicians. Then NYU's Tristan Buckmaster, who'd been pursuing that same approach with Anthropic's Levent Alpöge, cried foul. He has also alleged that OpenAI tried to bribe and threaten him. That's his allegation. Now the experts Scientific American spoke with say OpenAI's proof takes a genuinely different route, and the resemblance is only skin-deep. So the lab that announced the result described it wrong. Howlett's point is that OpenAI didn't understand its own model's proof, and a month later nobody fully does. The one safeguard is that the agents encoded their arguments in Lean, so a machine can check each step even when people can't follow the whole thing. And last week OpenAI released roughly four hundred more results. Howlett writes that there aren't enough experts on Earth to keep up. So the bottleneck isn't the model anymore. It's the reviewers. It's the same lesson as the coding-agent study we covered October 10th. Producing output is cheap now. The work piles up at review.
If the Maia racks and the Osaka megawatts were your favorite part today, try our sister show, The Data Center Daily. It's a daily briefing on AI compute: hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. Search for The Data Center Daily in your podcast app.
Two dates to watch: Argon's introductory pricing ends November 30th, and Anthropic and NTT Data are reportedly aiming to sign the Osaka deal by March 31st. The show notes link all four of today's sources. We'll be back tomorrow with the next AI Daily Briefing, a Lantern Podcast.