OpenAI Built Its First Chip in Nine Months, and It Just Beat Nvidia's Best Server on the Only Number That Matters

OpenAI spent nine months building a chip from a blank sheet of paper, and it just beat Nvidia’s best inference server on the number that actually decides who wins the AI race: work done per watt. Amazon is closing the marketplace that trained an entire generation of machine learning models, twenty one years after it opened. Google is sending AI agents into some of the most conservative law firms in the world. SoftBank is preparing one of the biggest corporate bond sales of the year just to keep its OpenAI bet funded. And in India, companies raised money on the stock market this August faster than in any month on record.

OpenAI’s homemade chip just beat Nvidia at its own game

OpenAI and Broadcom designed an inference chip called Jalapeño in nine months, from first sketch to a finished piece of silicon manufactured by TSMC, with Celestica building the racks around it. This week the first independent benchmarks came out, run by a group called InferenceX, and they are not close.

Jalapeño delivered 1.5 to 1.9 times more AI work per kilowatt than Nvidia’s GB300 systems across three large open models, and its latency was up to 3.6 times lower. For a company that has spent years paying Nvidia a premium for every chip it can get its hands on, building something that beats Nvidia’s current flagship on efficiency is a genuinely big deal.

There is a catch worth stating plainly, because it is the kind of detail that gets lost in a headline. Jalapeño uses newer HBM4 memory, while the GB300 it was benchmarked against uses the previous generation. Nvidia’s upcoming Rubin platform will use HBM4 too, and nobody has benchmarked Jalapeño against that yet. So this is a real result, not a fair fight with next year’s Nvidia. Jalapeño is also aimed only at low volume production later this year, not a mass replacement for Nvidia hardware anytime soon. What it proves is narrower and still significant: a company that is not a chipmaker can now design something that competes with the best chipmaker in the world has to offer, on the metric that determines how much a data center actually costs to run.

Amazon is closing the platform that trained the machine learning era

Amazon will shut down Mechanical Turk on September 30, ending a marketplace Jeff Bezos once described as “artificial artificial intelligence.” Launched in 2005, MTurk let businesses post small digital jobs, labeling images, transcribing audio, answering surveys, to a crowd of paid human workers. For close to two decades, that human labor was the invisible layer underneath a huge amount of machine learning research, including the image datasets that early computer vision models were trained on.

Amazon is not just retiring one product. SageMaker Ground Truth and Amazon Augmented AI, the company’s newer data labeling tools, are closing on the same date. A newer generation of startups, Scale AI, Mercor and Prolific among them, has taken over the business of recruiting people to train AI models, often paying more and offering more specialized work than MTurk’s flat per task rates ever did.

There is something fitting about the timing. The tool that used to be the punchline, “it’s not really AI, it’s just people behind a curtain,” is closing in the same year that a company can design a chip that beats Nvidia’s benchmarks. The scaffolding got kicked away because the buildings it held up don’t need it anymore.

Google is sending AI agents into corporate law firms

Google Cloud launched Gemini Enterprise for Legal this week, a version of its AI agent platform built specifically for law firms and corporate legal departments. Cleary Gottlieb, Freshfields, Weil and Williams & Connolly are named as launch partners.

The pitch is narrow and practical rather than flashy. The system connects into the document management and e-discovery software firms already run, and it is built to keep a firm’s existing ethical walls, the internal rules that stop one team from seeing another client’s confidential files, intact automatically rather than relying on lawyers to enforce them by hand. Contract review, due diligence document review and regulatory monitoring are the first jobs it is aimed at, and its legal research answers are meant to be grounded in primary law rather than whatever the underlying model happened to learn during training.

Google says a financial services version of the same idea has launched in parallel, with healthcare and life sciences versions coming later. Big Law has been one of the more cautious industries about handing real work to AI, given the cost of a hallucinated citation in a court filing. This is Google betting that cautious is exactly where the money is.

SoftBank is borrowing big to keep funding OpenAI

SoftBank is talking to banks about selling 10 billion to 20 billion dollars in bonds, priced in dollars and euros, as early as September. The money would refinance part of a 40 billion dollar bridge loan the company took out earlier this year to fund its stake in OpenAI.

Masayoshi Son’s company plans to put close to 65 billion dollars into OpenAI by October. At the top end, a 20 billion dollar bond sale would be the largest by any Asian company this year. SoftBank is also reportedly considering a 144A structure, a format that would let it sell the notes to a much wider pool of American institutional investors, something it has not done in more than a decade.

None of this changes what OpenAI is doing day to day. What it shows is how much of the capital behind the AI boom is now running through corporate debt markets rather than venture capital checks, with the risk sitting on SoftBank’s balance sheet rather than a fund’s.

India just had its best month ever for new stock sales

India’s stock market is on track for its strongest month on record for pricing new share sales, with close to 10 billion dollars in equity deals lined up in August. The single biggest piece is the government’s own 3.2 billion dollar sale of shares in Life Insurance Corporation of India, the country’s largest insurer. Manipal Health Enterprises added a 958 million dollar initial public offering, alongside a wave of smaller block trades and institutional placements.

What makes the number notable is what it is happening against. India’s benchmark Nifty 50 index has actually fallen so far this month, making this a rush of money into new issues rather than a market riding a broad rally. Analysts point to India’s own mutual funds and insurers, now large and well funded enough to absorb big deals on their own, plus retail investors and returning foreign funds, as the reason companies and the government felt confident enough to price so much stock at once.

A record month for raising capital is not a headline that jumps off the page the way a chess win or a rocket launch does. But it is the kind of number that tells you something is working underneath the surface: enough domestic savings, enough investor confidence, and enough demand for new Indian shares that a government insurer and a hospital chain could both raise money in the same month without spooking the market.

Every story here is about who is willing to bet money on what happens next. OpenAI and Broadcom bet nine months of engineering time that they could out design the biggest chipmaker on earth, and the early numbers say they were right. Amazon is betting that the labor market AI itself created has moved on for good. SoftBank is betting tens of billions of dollars, borrowed at scale, that OpenAI is worth it. And in Mumbai, a government insurer and a wave of Indian companies just bet that investors, domestic and foreign alike, are ready to put their money where that confidence is. So far, in every case, the bet is paying off.