← kaysauer.com AI & Tech · 30 August 2026

Both Ends of the Moat

Week 35 — two moves in the same week pull the AI compute stack in opposite directions: a customer ships silicon that beats NVIDIA on efficiency, and NVIDIA buys the hub every open model is distributed through.
This Week's Overview

This week the interesting action is one layer down from model scaling, in who owns the hardware and the distribution. Where last week was about which dial to turn next — params, data, or post-training — this week two moves pull the compute stack in opposite directions. OpenAI announced Jalapeño, its first inference chip, and the numbers are not a first-generation hedge: SemiAnalysis, invited to benchmark it in OpenAI’s own lab, says it beats Blackwell on performance-per-watt across almost every scenario. A company that was purely an NVIDIA customer 18 months ago now has silicon that’s competitive at the top. And NVIDIA is buying Hugging Face for a reported $13 billion — roughly 80× revenue — folding the default, vendor-neutral hub for open models into the dominant AI hardware company. Put together: NVIDIA’s moat is being chipped at from below in silicon and, apparently in response, extended upward into software and distribution. Also This Week covers the rest — the other Hot Chips announcements, Chinese open models still shipping weekly, AGI talk acquiring a date, the build-your-own-agent trend, and platform risk showing up in real API cutoffs.

Contents

Focus Topics

  1. NVIDIA buys the open-model hub
  2. OpenAI’s first chip beats Blackwell on efficiency

Also This Week

  1. The rest of Hot Chips
  2. Chinese open models keep shipping
  3. AGI talk acquires a date
  4. The center of gravity moves to engineering
  5. Build your own agent, or redesign for agents
  6. RL, and the weekly paper feed
  7. Benchmarking agentic inference
  8. Platform risk, from Cursor to neoclouds
  9. Can you read a model’s private thoughts?
  10. Foundation models for physics, not just language

Test Yourself

Read the newsletter and you'll be able to answer these.
  1. Roughly what multiple of Hugging Face’s annual recurring revenue is NVIDIA’s reported $13 billion price?
    Show answer
    B — roughly 80× Hugging Face’s ~$150M ARR, and nearly double NVIDIA’s own initial $7B offer back in January 2026. Hugging Face had doubled its customer base over the course of 2026.
  2. Why does folding Hugging Face into NVIDIA raise a neutrality concern that a normal acquisition wouldn’t?
    Show answer
    Hugging Face is the default, vendor-neutral place essentially every open model is hosted, benchmarked and downloaded from — the plumbing of the open ecosystem. Putting it inside the company that also sells most of the hardware those models run on gives one vendor a view into, and a hand on, the distribution layer for the whole field.
  3. OpenAI went from hiring the Jalapeño team to manufacturing tape-out in about how long, and why is that notable for a first-generation chip?
    Show answer
    About 16 months — an extremely fast ASIC cycle. First-generation chips are usually not competitive; SemiAnalysis’ point is that Jalapeño bucks that trend and beats every NVIDIA, AMD and Google chip they’ve tested on several open-source models, which they attribute to aggressive hardware–software co-design (and to AI genuinely being used to accelerate the chip design itself).
  4. What is Jalapeño’s headline result against NVIDIA’s Blackwell, and what handicap did OpenAI’s chip carry in that comparison?
    Show answer
    It beats Blackwell on performance-per-watt (token throughput per all-in-utility megawatt) across almost all scenarios — both low-latency and high-throughput — while running without multi-token prediction, whereas the competing chips were shown in their best configs with MTP on. OpenAI’s own figures: 1.5–1.9× more work per watt, 1.7–3.6× lower latency, 2.1–4.1× more performance on highly interactive workloads.
  5. SemiAnalysis calls the Blackwell comparison “incomplete and unfair.” What do they say the fairer comparison is, and why does timing matter?
    Show answer
    Jalapeño really competes with NVIDIA’s Rubin generation, which also uses HBM4. And Vera Rubin systems are shipping to customers now, while OpenAI won’t deploy Jalapeño into its own infrastructure until year-end — so the head-to-head is between a shipping product and a lab demo. Packaging and foundry (TSMC/CoWoS) capacity is also still a hard bottleneck for OpenAI.
  6. What was the “second-order story” about how Jalapeño’s low-level software got written?
    Show answer
    OpenAI says GPT-Astra plus Codex helped write and optimize the low-level kernels — bringing three previously unplanned open-weight models to high performance on the chip in about two months, with some attention and MoE kernels running 1.5–1.8× faster than existing human-expert-written code. Compiler and kernel work is being folded into the model-improvement loop.
  7. What did SemiAnalysis not get to test on Jalapeño, and why does that gap matter?
    Show answer
    They verified the shorter InferenceX runs in person but did not run the full suite or their AgentX benchmark — the long-context, multi-turn workload that stresses routers, prefix caches and offload infrastructure the way real production traffic does. A chip that looks great on single-turn 8k-in/1k-out can look worse under realistic agentic load.
  8. The two focus stories pull the compute stack in opposite directions. Describe each direction.
    Show answer
    Downward pressure on NVIDIA: OpenAI, formerly just a customer, now has an inference chip that beats NVIDIA on efficiency — frontier labs may no longer be strictly downstream of NVIDIA for inference economics. Upward extension by NVIDIA: buying Hugging Face moves NVIDIA from selling hardware into owning the software and distribution layer where models are hosted and discovered.
  9. According to the code-verification piece, what does DORA and related data show about AI coding tools?
    Show answer
    The velocity gains from AI coding assistants aren’t translating into real delivered productivity, because the volume of AI-generated code is outpacing teams’ ability to verify it — the bottleneck has moved to the verification layer, which most orgs haven’t tooled for.
  10. What’s the shared thread between OpenAI cutting off Cursor’s API access and the SemiAnalysis piece on neocloud security?
    Show answer
    Both are about counterparty risk in a fast-consolidating supply chain: an app-layer company can lose access overnight when its model provider also ships a competing product, and a lab spreading training and inference across many new GPU vendors inherits each one’s weak multi-tenant security (container escapes, kernel bypass, exposed dashboards).

From Last Week's Issue — 2026-08-23

  1. According to Jie Tang, what does a raw parameter count need to be considered alongside to actually mean something?
    Show answer
    How much data you have, where you intend to spend your compute, and who will run the model under what conditions. Scaling is a control panel of several dials — model size, pre-training data, compute per pass, mid-training, post-training-plus-RL — not one volume knob.
  2. What specifically produced GLM 5.3’s capability gains, given identical architecture and size to GLM 5.2?
    Show answer
    C — simulated multi-day engineering tasks with real codebases and experiment results that the model works through end to end. Gains from the post-training-plus-RL dial alone, model size untouched.
  3. What does Glean’s Waldo system do before a query ever reaches an expensive frontier model?
    Show answer
    It acts as a cheap agentic-search model that reads the query first, decides how to break it down and which tools it needs, and assembles the raw materials — so the expensive model is only invoked once the groundwork is done.
  4. What’s the “written-down vs computed knowledge” distinction from the Waymo-vs-Tesla piece?
    Show answer
    Written-down knowledge is surveyed, validated and inspectable ahead of time (Waymo’s structured object lists and roadgraph); computed knowledge is inferred fresh, more flexible but opaque (Tesla’s camera-only neural network). The same tension runs under the scaling and simulation stories.

Focus Topics

NVIDIA buys the open-model hub

[AINews] NVIDIA buys HuggingFace for $13B, as OpenAI publishes their HF incident retro — Latent Space, 27 August 2026

After first offering roughly $7 billion back in January, NVIDIA is now reported — per The Information’s scoop and follow-up confirmation — to be acquiring Hugging Face for about $13 billion. That is on the order of 80× Hugging Face’s ~$150 million in annual recurring revenue, a multiple that only makes sense if you’re buying strategic position rather than a cash-flow business. Hugging Face had roughly doubled its customer base over 2026, but the price is really about what the company is: the default, vendor-neutral hub where nearly every open model gets hosted, versioned, benchmarked and pulled from.

That is what makes this different from NVIDIA’s other recent moves in the open ecosystem. Buying a startup’s compute contracts or funding open-model training expands demand for GPUs without touching anyone’s independence. Owning the Hub is owning a chokepoint — the piece of infrastructure that every lab, cloud and hobbyist routes through regardless of whose hardware they run on. Latent Space’s own framing was cheerful (“we love it when the good guys win”), and Hugging Face has been a genuinely good steward of open source. But the structural question stands: the dominant seller of AI hardware would now also run the distribution layer for the models, with the visibility into usage and the standard-setting power that comes with it.

The timing is not incidental. The deal landed in the same week that Z.ai’s GLM-5.3-Flash and a new Qwen Flash model — both strong, both open, both trained on Chinese accelerators — had the post-Hot-Chips conversation focused on whether “Western open AI” has a coherent home. The same AI News issue also carried OpenAI’s postmortem on the Hugging Face security incident from late July, a reminder that the Hub is critical enough infrastructure that its failures are industry events.

Why it matters

If this closes, the open-model ecosystem’s main piece of neutral infrastructure will sit inside NVIDIA. Pair it with Jalapeño — a major customer building competitive silicon — and the picture is NVIDIA’s position being pressured in hardware and, seemingly in response, extended into software and distribution. Worth watching for regulatory friction and for how the labs that depend on the Hub react.

OpenAI’s first chip beats Blackwell on efficiency

OpenAI Jalapeño: Better Than Nvidia Blackwell — SemiAnalysis, 25 August 2026

OpenAI used the Hot Chips conference to unveil Jalapeño, the inference chip it has been building with Broadcom since mid-2024. The striking part isn’t that it exists — the Broadcom partnership was announced in June — it’s the performance. SemiAnalysis was invited into OpenAI’s lab to run its InferenceX benchmark on the chip, and their verdict is blunt: Jalapeño beats every NVIDIA, AMD and Google chip they’ve tested on several open-source models. First-generation accelerators are almost never competitive; this one leads. They credit aggressive hardware–software co-design, a roughly 16-month path from team hiring to tape-out, and — notably — AI genuinely being used to speed up the chip design.

The headline number is performance-per-watt: token throughput per all-in-utility megawatt, where Jalapeño “smokes every other chip,” and does it without multi-token prediction while the competing chips are shown in their best MTP configurations. It wins in both low-latency and high-throughput regimes rather than being tuned for one point on the curve — over 700 tokens/sec/user at concurrency 1 on DeepSeek R1 with plain single-token decoding, no speculative decoding, no prefill/decode disaggregation. OpenAI’s own figures against NVIDIA GB200/GB300 systems: 1.5–1.9× more work per watt, 1.7–3.6× lower end-to-end latency, and 2.1–4.1× more performance on highly interactive workloads. A separate detail worth its own headline: GPT-Astra and Codex helped write the low-level kernels, bringing three unplanned open-weight models to high performance in about two months, some kernels running 1.5–1.8× faster than expert-written code.

SemiAnalysis is careful about the caveats. Every number came from OpenAI; they verified the InferenceX runs in person but not the full suite, and didn’t see AgentX results — the long-context, multi-turn benchmark that reflects real production cache behavior, where a chip that shines on single-turn tests can stumble. They also call the Blackwell comparison “incomplete and unfair”: Jalapeño’s real competition is NVIDIA’s HBM4-based Rubin generation, and Vera Rubin systems are shipping to customers now while OpenAI won’t put Jalapeño into its own datacenters until year-end. Packaging and foundry capacity (TSMC, CoWoS) remain a hard limit on how fast OpenAI can scale it.

Why it matters

The concrete claim — verified in person by a skeptical third party — is that a company that was a pure NVIDIA customer 18 months ago now has inference silicon competitive with the best. Even with the deployment lag and the Rubin caveat, that’s the strongest evidence yet that frontier labs won’t stay strictly downstream of NVIDIA on inference economics. It’s also the clearest example so far of AI-assisted chip and kernel design producing a real, shipped result.

Also This Week

The rest of Hot Chips

Jalapeño dominated the coverage, but the 37th Hot Chips also brought Cerebras’s CS-5, Groq’s 3 LPX, and Apple’s M6 — a snapshot of a custom-silicon race that now has frontier AI labs, wafer-scale specialists, and consumer-chip makers all pushing on inference efficiency at once. The through-line matches the focus stories: the interesting competition has moved from “who trains the best model” toward “who runs it cheapest per watt.”

Chinese open models keep shipping

This week’s agent-and-model roundup led with GLM-5.3-Flash (320B total / 18B active params, 1M context, MIT license, trained entirely on Chinese accelerators, scoring 57 on the Artificial Analysis Intelligence Index at ~$0.09/task), plus a Qwen3.8-Flash, an Hy4 preview, Claude’s new built-in browser, and Terminal-Bench-Science. The cadence — a capable open Flash-tier model roughly every week — is the backdrop to the Hugging Face deal.

AGI talk acquires a date

A roundup covering reports that OpenAI is targeting an internal “AGI bar” by the end of 2026 — less interesting as a prediction than as a marker of how openly frontier labs are now attaching public timelines to the term, and what that does to fundraising and expectations.

The center of gravity moves to engineering

Andrew Ng is repositioning DeepLearning.ai around “AI Engineering,” backed by an analysis of 10,000+ job postings — a mainstream signal that the work is shifting from training models to building reliable systems around them. Two companion pieces on what that work actually involves: why the verification layer is now the bottleneck for AI-generated code, and a walkthrough of speculative decoding, the technique behind roughly 3× faster inference with unchanged outputs.

Build your own agent, or redesign for agents

Ramp built its own in-house coding agent, Inspect, rather than adopting a frontier lab’s — and says the bet is paying off. Lovable’s CTO argues the future of SaaS is apps redesigned as MCP-powered “capabilities” that agents, not just humans, can use. And Meta reportedly considered cutting engineering teams by 60% out of fear of leaner AI-native startups — a look at how AI anxiety is reshaping org design at the top.

RL, and the weekly paper feed

Cameron Wolfe’s from-first-principles guide to reinforcement learning for LLMs traces the arc from early alignment work to today’s reasoning- and agent-focused frontier — a useful single reference given how central RL has become to capability gains (see last week’s GLM 5.3 story). The recurring research digests also landed: top papers for Aug 17–23 and Aug 24–30.

Benchmarking agentic inference

SemiAnalysis released AgentX 1.0, an open-source (Apache 2.0) multi-turn agentic-coding inference benchmark at 1M context — the first real open benchmark for the long-context, multi-turn workloads that have overtaken fixed-length single-turn traffic as the dominant production pattern. It’s also, per the Jalapeño piece, the suite OpenAI’s chip hasn’t yet been shown running.

Platform risk, from Cursor to neoclouds

OpenAI cut Cursor’s API access amid the Elon-vs-Altman fallout — a concrete case of an app-layer company’s dependence on a model provider that also ships a competing product. Further down the stack, SemiAnalysis argues most GPU neoclouds have weak multi-tenant security — container escapes, kernel bypass, exposed Grafana — a systemic risk as the biggest labs spread workloads across many new vendors (the piece includes a ClusterMAX 3.0 preview).

Can you read a model’s private thoughts?

A walkthrough of whether the encrypted “private” reasoning traces that Anthropic, OpenAI and Google return to API clients can actually be reconstructed by an attacker — the applied version of the reasoning-trace extraction vulnerability from two weeks ago, where a signed reasoning block is replayed into a weaker same-provider model that’s prompted to transcribe it.

Foundation models for physics, not just language

An interview with Anima Anandkumar on why physical systems — weather, fusion reactors — still lack the kind of general-purpose foundation models language has, and her argument for what building them would take. A useful counterweight to a week otherwise dominated by inference economics and M&A.