In partnership with
In Today’s Issue:
🌶️ OpenAI's Jalapeño chip outruns Nvidia's flagships in its first public results
🤖 XPeng's robotics unit raises $900M at a $6.3B valuation
💰 Anthropic pitches IPO investors a $30 trillion market
📱 Figure is building the world's biggest robot dataset from filmed chores
✨ And more AI goodness…
⚡ The Signal
The AI giants are done renting their futures: they are building the whole stack themselves, and pricing it in trillions.
OpenAI published the first measured results for Jalapeño, its own inference chip, and they challenge Nvidia's best. XPeng raised over $900 million to build humanoids on chips, models, and factories it owns end to end. Figure came out of stealth with an app that pays people to film everyday tasks, because the training data its robots need cannot be bought. And Anthropic is preparing to tell IPO investors that the prize for all this is a market above $30 trillion. Control of the stack, from silicon to data, is becoming the moat; whether the revenue follows is the question Wall Street now has to price.
All the best,
Kim Isenberg
(XPENG)
🤖 XPeng's Robot Unit Banks $900 Million
XPeng's robotics unit has raised over $900 million at a valuation above $6.3 billion, the largest single private round in China's embodied-AI industry. IDG Capital led the round, with Tencent and Alibaba joining as strategic investors. The money goes into IRON, XPeng's general-purpose humanoid that runs the company's Physical AI model on three in-house Turing chips (2,250 TOPS of on-board compute) and moves on 76 degrees of freedom; mass production is planned for the end of 2026, with deliveries in China and overseas from 2027.
👉 tl;dr: China's biggest embodied-AI round is a bet that an EV maker's factories can put humanoids into mass production first.
(Jason Henry for WSJ)
💰 Anthropic Pitches a $30 Trillion Market
Anthropic is expected to tell IPO investors it sees more than $30 trillion in potential revenue, the Wall Street Journal reports, topping SpaceX's $28.5 trillion record. The figure is a total addressable market (TAM) estimate, the revenue a company could capture if it owned its entire market; Anthropic reportedly gets there by counting the full scope of work AI models could complete. The company is said to be targeting a valuation of about $2 trillion and a raise of up to $100 billion, with a listing as soon as September. For scale: the 191 tech companies in the S&P 1500 booked $2.4 trillion in revenue last year combined.
👉 tl;dr: Wall Street is being asked to price AI against the value of all human work, and Anthropic just set the high bar.
(Figure)
📱 Figure Pays People to Film Their Chores
Figure came out of stealth with Index, an app that pays people to film everyday tasks and feeds the footage to Helix, its robot brain. In four months the app has logged 264,000 downloads across 108 countries, with over 44,000 weekly active users uploading 16 million videos, 30 minutes of footage arriving every second. Figure has paid its Creators $15 million so far and says it will spend over $1 billion on data and compute in the next 12 months, calling Index the largest and most diverse physical-AI dataset in the world.
👉 tl;dr: Figure thinks the scarcest resource in robotics is real-world data, and it is buying it straight from your kitchen.
🎬 Watch This
What happens once AI can automate AI research?
Dwarkesh Patel spends more than two hours with Ryan Greenblatt, chief scientist at the AI-safety lab Redwood Research, on exactly that question: how a model goes from junior engineer to running thousands of automated researchers, why that feedback loop could compress a decade of progress into a couple of years, and where it could break. Clips from the episode have been rolling out all week.
"I've been working towards AGI my whole life and now, like many of you, I feel it is close at hand."
– Demis Hassabis, DeepMind co-founder, announcing his move to Chief Scientist of Alphabet in August
Shopify CEO Tobi Lütke is threatening to ban Claude Code at Shopify unless Anthropic supports the tool-neutral AGENTS.md standard instead of insisting on its own CLAUDE.md, which he says causes "split brain problems" for teams mixing agents. An Anthropic engineer has already replied under the post: support is reportedly on the way.
The Chip OpenAI Built for Itself Just Outran Nvidia
The Takeaway
👉 Jalapeño, OpenAI's first custom inference chip, delivered 1.5 to 1.9x more AI work per watt and 1.7 to 3.6x lower latency than Nvidia's GB200/GB300 systems on the public InferenceX benchmark.
👉 The lead held across three open models (GPT-OSS 120B, DeepSeek R1, Kimi K2.5 1T); OpenAI says it widens further on its own frontier models.
👉 AI built its own hardware: earlier GPT generations helped design and bring up the chip, and OpenAI's latest models now optimize and program it.
👉 The numbers are OpenAI's own runs; SemiAnalysis verified only some of them on-site, and the chip still has to clear production qualification before the ramp.
OpenAI just showed receipts for its independence from Nvidia. At the Hot Chips conference on August 25, the company published the first measured results for Jalapeño, its first custom inference chip, tested on InferenceX, a public benchmark from SemiAnalysis that measures the full process of serving an AI request. Across GPT-OSS 120B, DeepSeek R1, and the trillion-parameter Kimi K2.5, Jalapeño delivered 1.5 to 1.9 times more AI work per watt and 1.7 to 3.6 times lower end-to-end latency than Nvidia's flagship GB200 and GB300 systems; on highly interactive workloads the gap grew to 2.1 to 4.1 times.
OpenAI's chart: throughput per kilowatt at the fastest decode speed Nvidia's systems reach (OpenAI via The Decoder)
The design explains the numbers. Jalapeño was built for one job, serving language models and especially agents, so OpenAI co-designed the chip, memory, network, and serving software around real workloads: the model's working memory stays local instead of shuttling between chips, and the same silicon handles both reading a prompt (compute-hungry) and writing the answer token by token (memory-hungry) without trading one off against the other. The stranger part is who built it: OpenAI says earlier GPT generations helped design and bring up the chip, while its latest models now optimize and program it.
The Jalapeño inference chip (OpenAI)
The caveats deserve their own sentence. These are OpenAI's own benchmark runs: SemiAnalysis verified some of them on-site, but the claim that the advantage widens on frontier models comes from internal testing nobody outside can check. And a benchmark win is not a deployed fleet: the chip is still in production qualification, with the ramp promised "in the months ahead".
Why it matters: Inference cost per watt is becoming the margin that decides who can afford to serve AI agents at scale; if these numbers survive deployment, OpenAI has a credible in-house alternative to the supplier almost every AI lab pays.
Sources:
🔗 https://openai.com/index/jalapeno-first-results/
🔗 https://the-decoder.com/openais-first-custom-chip-jalapeno-reportedly-beats-nvidias-blackwell-and-rubin-in-inference-benchmarks/
🔗 https://techcrunch.com/2026/08/25/openais-jalapeno-chip-is-built-for-fast-inference-at-scale-benchmarks-show/
Cut Lead Review From Hours To Minutes
Sign up for a free trial of Attio, the agentic CRM.
Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.
Review your pipeline in Claude, synced live from Attio via MCP.
That's it.
(OpenAI, InferenceX benchmark)
The chart: OpenAI's peak-efficiency chart from the Jalapeño results: mixed tokens per second per kilowatt of power, at each system's best operating point. Running GPT-OSS 120B, Jalapeño moves 85,448 tokens per second per kilowatt against 44,960 for the best system on the market today, a 1.9x gap; on DeepSeek R1 (1.7x) and the trillion-parameter Kimi K2.5 (1.5x) the lead is smaller but holds.
The lesson: OpenAI says outright that work per unit of power is the number that matters: today's data centers are capped by the electricity they can draw long before they run out of chips. If a watt of Jalapeño does up to nearly twice the work of a watt of Nvidia's best, the same grid connection serves far more users.
The caveat: Every bar is OpenAI's own run, normalized by rated chip power, and "Existing best" means the best commercially available systems OpenAI tested, Nvidia's GB200 and GB300, not Nvidia's next generation.
🔬 Faster Scientists, Shallower Science?
⚡ Bottom line: A new model by four US researchers finds AI time savings push scientists to start more projects and polish each one less.
💡 Why it matters: In two of three modeled scenarios, paper quality drops even when the AI itself works perfectly.
🔎 What it means: Aimed at deep work, AI improves science; aimed at speed, it mostly fills journals faster.
The claim in one sentence: even a flawless AI assistant could make published science worse, because the time it saves gets spent starting new projects instead of finishing old ones properly. That is the result of a mathematical model by Eamon Duede, Kevin Gross, M.J. Crockett and Carl Bergstrom, four researchers across Purdue, NC State, Princeton and the University of Washington, posted to arXiv in July and picked up by The Decoder this week.
The logic is everyday economics. When a tool saves you time, your remaining hours become more valuable, so you spend them where the payoff is biggest: launching the next promising project rather than polishing the current one. The team borrowed the math biologists use for foraging animals, which leave a berry bush while it still has berries because the next bush pays better. In the model, researchers treat papers the same way.
The paper's model of a research project: a discovery phase, a development phase, and the option to go back to the drawing board (Duede et al., arXiv)
They ran three scenarios. When AI helps evaluate early ideas, researchers get pickier but give each surviving project less care. When it speeds up writing and formatting, weaker projects suddenly become worth publishing, and journals fill with more but shallower papers. Only when AI takes over the deep, voluntary work itself, the extra experiments and more careful analysis, does quality rise. Two paths out of three end in worse science. The authors put it bluntly: "As a labor-augmenting technology, LLMs increase the opportunity cost of our time, impelling us to do more, less well—rather than the same amount, better."
The paper, posted to arXiv on July 19 (arXiv)
The pattern is already visible outside the model: a METR study found experienced developers using AI tools were 19 percent slower while feeling 24 percent faster, and arXiv has tightened its rules against a flood of AI-assisted submissions. The authors' fix: aim the tools at the deep work, and reward thoroughness over throughput.
Winning Startups Aren't Bigger. They're Leaner.
Founders deploying AI agents across sales, marketing, and customer service are closing more with smaller teams. 65% more sales leads. Less headcount.
Download the free Practical Guide to Agentic GTM for Startups and start building the stack your competitors dream of.
