If you were planning to rent a big block of AWS capacity in 2027, Andy Jassy has news for you. Most of it is already gone. And a good chunk of 2028 is gone too. Good morning, it's Monday, October 26th, and you're listening to AI Daily Briefing. Today: Amazon's cloud is booked solid, CNBC looks at Nvidia's growing pile of loan guarantees as neocloud stocks slide, an independent leaderboard puts Moonshot's Kimi K3.1 Pro just behind the top three, and a recruiting company's AI lab ships an open thirty-three billion parameter agent model. Follow the show and the next briefing lands in your feed on its own.
Pragya Singha Roy at Mint, following Jaspreet Singh's report for Reuters:
Infrastructure is not just about data centres, Jassy said. Power supplies and memory parts are also in tight supply, and he does not foresee the memory crunch coming to an end in 2027.
The setting was a video podcast on Friday with Andreessen Horowitz's Ben Horowitz and Erik Torenberg. Jassy said AWS capacity is largely sold out through 2027, with a significant share of 2028 already committed. Some is held back for on-demand use, but the big chunks are spoken for. So when your account rep says capacity is tight, they're not bluffing. For a startup that means on-demand is what's left, and on-demand is the slice Amazon has the least of. Amazon had guided to about two hundred twenty billion dollars in capital spending for 2026. Per Reuters, Jassy said it could invest a lot more and that it wouldn't be a bad use of capital. The constraint he named isn't only buildings. Power, high-bandwidth memory and ordinary DRAM are all short, and he doesn't see the memory squeeze ending in 2027. He also made the Trainium pitch: thirty to forty percent better price-performance than comparable alternatives. That's Amazon's number. What I'd take from it is that Jassy said inference is still an order of magnitude too expensive for AI to be as everywhere as it could be. Coming from the man selling the inference, that's quite an admission. Amazon reports third-quarter results this Thursday, October 29th.
Kif Leswing and Isabel O'Brien at CNBC, writing Saturday:
As of its last quarterly financial filing in July, Nvidia had $36 billion in off-balance-sheet AI cloud agreements, or backstops, for neoclouds, in addition to its guarantee on up to $105 billion for chips, land and power at a proposed 10 gigawatt OpenAI data center in Ohio. The company signed as a tenant for $20 billion in future data center leases, which will later be assigned to third parties.
That's Nvidia acting as the credit rating behind other companies' data centers. In most of these deals it pays nothing unless someone defaults. CNBC's point is that the market has started pricing in the chance that someone does. The numbers for October: CoreWeave, Nebius and IREN are each down at least thirty percent this month. The spread on CoreWeave's credit default swaps is past eleven percent, and yields on its 2031 and 2032 bonds topped fifteen. Nvidia's own stock is down seven point six percent for the month. Fifteen percent on a bond for a company whose business is renting out GPUs. That's late-stage 2021 territory. Some continuity. Last Monday we covered Microsoft walking away from Nscale's Loughton site. CNBC notes Nvidia is an investor in Nscale, which filed to go public in the U.S. in September and hasn't updated those plans in weeks. Firmus, another Nvidia-backed cloud, pulled its IPO this month. And on Thursday we covered the two hundred seventy-nine billion in supplier commitments. Nvidia is carrying risk at both ends of its supply chain. Jensen Huang's answer, given to CNBC in Berlin last Thursday: demand for neoclouds is off the charts, and what's missing is working capital. Which happens to be exactly what Nvidia is supplying. Moody's offers the calmer view. David Gonzales says chips will likely hold values above their contractual residual amounts in the near term, which would limit Nvidia's exposure.
Artificial Analysis, publishing its evaluation Friday, with a write-up from the OfficeChai team:
Scoring 55 on the index, K3.1 Pro trails only three models — Claude Opus 5.5 (max with fallback, 58), Claude Sonnet 5.5 (max with fallback, 56) and Muse Wonder Full (max, 56). It outscores OpenAI’s GPT-6 Astra (53), Google’s Gemini 4 Argon (53), and every other model currently on the board, including Meta’s Muse Wonder (max, 52) and GPT-6.1 Sol (52).
That's eleven points over Kimi K3 at the exact same size: two point eight trillion total parameters, a hundred four billion active. What matters is who did the measuring. This isn't Moonshot's own table. It's a third party running the same ten evaluations on everyone. The biggest jump is in agentic work: sixty-seven percent on Terminal-Bench 4.0, the top score of any model, up from thirteen for K3. GDPval-AA, which tests real-world work tasks, is up three hundred twelve Elo, good for third place. And the price is the part that gets my attention. Moonshot cut its API pricing by a third, to two dollars per million input tokens and ten per million output. It uses about fourteen percent more output tokens per task, so cost per task on the index comes out flat: a dollar ninety-eight against two dollars. Near-frontier agent scores at that price change what I'd route where. The weights aren't out yet. Artificial Analysis says that once they are, it becomes the top open-weights model, and the gap to the best closed model shrinks from twelve points to three. The BYOBot roundup puts the promised date at November 4th. Launch week had problems, too. Per the iyooo roundup, Kimi has apologized for rate limits and unexpected credit drain on K3.1 Pro, reset usage for every subscriber and added a quarter of a month's quota as compensation. AWS isn't the only one short on capacity.
Nanbeige LLM Lab, on its Hugging Face model card for Nanbeige4.3-33B:
We compare Nanbeige4.3-33B with recently released frontier models at around the 30B-parameter scale: Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B. All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.
A thirty-three billion parameter dense model, no mixture of experts, under Apache 2.0, with a five hundred twelve thousand token context. Asif Razzaq at MarkTechPost notes the lab sits inside BOSS Zhipin, the recruitment platform, and puts the BF16 weights at about sixty-six gigabytes. The lab calls it a teacher model. The plan is to distill its agent skills into smaller models, starting with a three billion parameter version. In its own table it leads seventeen of twenty rows, but trails Qwen on SWE-Bench Pro and scores just twenty-three point eight on Terminal-Bench 4.0. Put that next to Kimi's sixty-seven on the same benchmark and you can see what a hundred billion active parameters buys you. But the row I keep looking at is Muse Glimmer. Meta's own report, which we covered last Wednesday, had Glimmer at fifty-one point two on SWE-Bench Pro. In Nanbeige's harness, it gets thirty-six point nine. Treat that carefully. It's a rival lab, using its own harness, mini-swe-agent, and those are its own numbers. But it's the first time we've had someone outside Meta run Glimmer against the field, and it doesn't flatter it. One commenter on the Digg thread who's used past Nanbeige models expects the results to mostly hold, but warns it may be fragile and prone to doom loops. That's the failure I'd test first. A model that gets stuck repeating step seven isn't an agent.
Two of today's four stories were really about megawatts and memory. The Data Center Daily covers that layer every day: hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. Find The Data Center Daily in the same app you're using right now.
That's Jassy's sold-out cloud, Nvidia's backstops, Kimi K3.1 Pro on an independent leaderboard, and Nanbeige's open teacher model. Every source is linked in the show notes. AI Daily Briefing is a Lantern Podcast. We'll be back tomorrow, Tuesday the 27th.