TechWithAdit (TWA)

OpenAI just unveiled its first ever custom AI chip — named Jalapeño, built with Broadcom — designed to run every ChatGPT conversation at roughly 50% of the cost of Nvidia GPUs. The chip was delivered to Sam Altman by Broadcom's CEO this morning. This is the moment OpenAI stopped renting intelligence and started building the factory that makes it.

While OpenAI makes its infrastructure move, the field's best minds are questioning whether chatbots are even the right destination. Fei-Fei Li, the "Godmother of AI," and a growing cohort of founders are publicly pivoting to "world models" — AI that can read a room, not just a book. Former Infosys CEO Vishal Sikka has just launched a startup to challenge the IT services world. And new SignalFire data says engineering jobs, far from disappearing, are the most resilient in the entire AI-era economy.

Fei-Fei Li, widely known as the "Godmother of AI" and the scientist whose ImageNet project launched the deep learning era, is now leading a wave of researchers and investors pivoting away from large language models toward what the field is calling "world models." The idea: AI cannot be truly intelligent if it can only read a book. It also needs to read the room. World models learn the statistical structure of space and time — how light falls on a surface, how objects respond to force, how a garden looks from an angle no camera has captured. Louis Castricato, a researcher who quit his PhD at Brown University after eight years studying LLMs, put it plainly: "We have basically passed the point of doing real fundamental LLM research. Now it's just applications." He has since started a company called Overworld.

The pivot is gaining serious investment traction. Li describes world models as "one of the most important and most overloaded terms in AI today." For scientists like Carnegie Mellon's Martial Hebert, who has spent four decades on robotics, world models are the faster and cheaper path to physical AI — the kind that can actually interact with the real world rather than describe it. The implication is significant: the companies dominating AI in 2026 are not necessarily the ones who will dominate physical AI by 2030.

The insights: Chatbots have captured the capital. World models are capturing the conviction. When the researchers who built LLM foundations start calling them a dead end, the next investment cycle in AI has probably already started — and it looks like robotics infrastructure, not another GPT.

SignalFire's latest labor market data tells a story the headlines have been getting wrong: engineers are not being displaced by AI — they are being hired in greater numbers. Software engineers are now making up a larger share of total new hires across tech companies than they were before the AI boom, not a smaller one. The pattern holds even at companies that have cited AI in their layoff announcements. What is being cut are non-technical roles. What is being added are engineers who can build, deploy, and maintain AI systems.

The fear that drove a generation of students away from software engineering degrees — the idea that AI would automate programmers out of existence — is being contradicted in real time by the actual hiring data. AI is changing what engineers do, not eliminating the need for them. The companies building with AI at scale are discovering they need more engineers than before, not fewer.

The insights: The AI displacement story has been running at the wrong resolution. It is accurate at the macro level — AI is eliminating jobs — but it is wrong about which ones. Operators who cut their engineering headcount to save costs in 2025 may be competing for the same talent pool in 2026 at significantly higher prices.

Vishal Sikka, who served as CTO of SAP and later as CEO of Infosys before departing under board pressure in 2017, has quietly launched a new startup called Hang Ten Systems, backed by Mayfield and Aramco Ventures. The company brings together a team of veterans from SAP, Infosys, and Sikka's previous AI venture VianAI. The thesis: the traditional IT services model — large teams doing managed work on a time-and-materials basis — is structurally vulnerable to AI-native disruption, and a well-funded startup with deep enterprise relationships and AI expertise can move into that gap faster than the incumbents can defend it.

Sikka's track record matters here. At Infosys, he ran one of the world's largest enterprise IT organizations and built its early AI platform. He left with deep institutional knowledge of exactly how these contracts work and where the inefficiencies live. Hang Ten Systems is not a software company building tools for IT firms — it is positioning itself as an AI-native alternative to them.

The insights: The threat to Indian IT services is not coming from silicon valley chatbots. It is coming from operators who understand the enterprise IT business from the inside and are rebuilding it with AI from scratch. Sikka knows those contracts better than most. That is a more credible threat than a benchmark score.

TechCrunch reports that Europe is pushing back hard against the US MATCH Act, which would extend semiconductor export controls to older-generation deep ultraviolet lithography tools — gear that ASML, the Dutch company that manufactures the world's most advanced chip-making equipment, first shipped roughly a decade ago. ASML CEO Christophe Fouquet has made the European position clear: these are the only machines China can currently buy from ASML, and restricting them would not meaningfully slow China's AI chip development — it would primarily damage ASML's revenue and European industry's competitiveness.

The pushback marks an increasingly visible fracture in the transatlantic tech alliance. The US wants a broader chip war. European governments and companies, whose economic interests are directly harmed by export controls they did not design and cannot veto, are now saying so publicly through their CEOs and trade bodies. The MATCH Act remains under debate in Congress, but European resistance has emerged as a meaningful political counterweight.

The insights: The chip war is no longer just a US-China story — it is becoming a US-Europe-China triangle. For any company in the semiconductor supply chain, the next twelve months of ASML access and export control negotiations will define the cost structure of AI compute for the rest of the decade.

OpenAI and Broadcom unveiled Jalapeño today — OpenAI's first custom AI accelerator, a chip designed from scratch to run large language model inference. The name is not accidental — it signals heat, speed, and a direct provocation. Engineering samples were delivered to CEO Sam Altman and President Greg Brockman by Broadcom CEO Hock Tan and President Charlie Kawwas at OpenAI's San Francisco headquarters this morning. The chip is already running GPT-5.3-Codex-Spark in the lab at production target frequency and power.

The economics are the headline. Broadcom CEO Hock Tan said in an interview that early testing shows cost savings of roughly 50% compared with typical AI GPUs. OpenAI's own language was more cautious — "substantially better performance per watt than current state-of-the-art" — but the direction is the same. Jalapeño is not a general-purpose accelerator. It is a blank-slate design for modern LLM inference: it reduces data movement, balances compute, memory, and networking to achieve utilization far closer to theoretical peak performance, and is optimized specifically for the interactive latency demands of ChatGPT, Codex, and agentic products at scale. Built in partnership with Broadcom and manufactured by TSMC, with server systems from Celestica, the chip will begin small prototype deployments by end of 2026 and full production ramp in 2027 and 2028 across Microsoft and other infrastructure partners.

What this does not mean: OpenAI is not ending its relationship with Nvidia. Jalapeño is an inference-only chip, and OpenAI remains one of Nvidia's largest training customers. But inference is where the bills pile up and where users feel the product — and that is exactly where OpenAI is now building its own alternative. The chip is described as step one of a multi-generation hardware roadmap. Broadcom's framing was expansive: "This is just the beginning of a multi-generation roadmap. By co-developing our industry-leading silicon directly with OpenAI, we are enabling the deployment of gigawatt-scale data centers with Microsoft and other partners beginning in 2026."

The insights: Every major cloud company eventually builds its own chips — Google has TPUs, Amazon has Trainium, Meta has invested years in custom silicon. OpenAI just joined that club. The immediate impact for Nvidia is limited: training demand remains intact and Jalapeño is still unproven at scale. The long-term signal is the one that matters. AI leadership is no longer purely about who has the best model. It is about who controls the economics of running that model at a billion conversations per day. OpenAI just made its first move toward controlling both.

Was this email forwarded to you? Don't miss any updates — Subscribe to TechWithAdit for sharp, no-noise tech intelligence. Stay sharp. — Adit