AI Daily Brief: OpenAI builds its own chip, a Chinese model undercuts everyone on price

Delegates are gathering in Geneva today for the UN's first Global Dialogue on AI Governance, but the more consequential news this week happened in fabs and training runs, not conference rooms. Three stories capture it: a lab that decided renting chips wasn't enough, a memory industry that can't make enough of them, and a model race that just got a lot cheaper to enter.

OpenAI stops renting chips and starts designing them

OpenAI and Broadcom unveiled Jalapeño, a custom inference chip built specifically for the way OpenAI's own models move data, and the numbers are the story: early testing points to roughly half the inference cost of current state-of-the-art hardware, with performance-per-watt gains to match. What's more striking is the timeline. Design to tape-out took nine months, reportedly one of the fastest ASIC development cycles ever attempted at this scale, partly because OpenAI used its own models to accelerate parts of the chip design process itself. Jalapeño ships as the first piece of a multi-generation compute platform targeting deployment by the end of the year, with Broadcom handling silicon implementation and Celestica building the racks around it. Every major lab has talked about the economics of inference becoming the real bottleneck in AI; OpenAI is the first to answer that by becoming a chip company on the side.

Micron and SanDisk are the quiet winners of the AI boom

While OpenAI designs chips, the memory industry is racing to keep up with everyone who already needs them. Micron's stock hit an all-time high after it signed a strategic agreement with Anthropic spanning HBM and DRAM supply, joint architecture design for AI workloads, and a strategic investment in Anthropic's $65 billion Series H round. It's a similar story one rung down the storage stack: SanDisk's data center business grew 645% year over year on a surge in enterprise SSD demand, and the company is shifting away from spot pricing toward multi-year supply agreements with prepayment commitments, the clearest sign yet that hyperscalers see this demand as structural rather than a spike. Memory chips used to be the boring, cyclical part of the AI trade. They're now arguably the tightest bottleneck in the entire stack.

The model race splits into a speed lane and a price lane

Two model releases this week point in opposite but related directions. Google's Gemini 3.5 Flash launched as the company's fastest agentic model yet, running roughly 4x faster on output tokens than other frontier models while leading on coding and long-horizon agent benchmarks, at under half the cost of flagship-tier models. Meanwhile Zhipu's GLM 5.2, a free, open-weight model out of China, has pulled to within a percentage point of Anthropic's Opus 4.8 on a closely watched agentic benchmark, at around a fifth of the cost, and developer traffic to it is climbing faster than DeepSeek's did after its own splashy debut in April. Read together, the two releases say the same thing from different directions: the ceiling on model capability is getting less interesting than the floor on what it costs to use one.

What it means for India

India has a stake in every thread above. Micron's assembly and test plant in Sanand, Gujarat, inaugurated by the Prime Minister in February, is meant to handle roughly 10% of Micron's global memory output by 2027, which means the same supercycle lifting Micron's stock is now partly running through Indian soil. The Semiconductor Mission 2.0, folded into this year's budget, has already cleared twelve fab and packaging projects worth roughly ₹1.64 lakh crore, with an explicit pivot toward memory and AI-specific chip design rather than general-purpose silicon. On the governance side, Minister of State Kirti Vardhan Singh is leading India's delegation at the Geneva dialogue this week, one of the largest blocs at a summit built specifically to keep AI governance from being written solely by the countries that already have the compute. And a cheap, capable open model like GLM 5.2 is arguably better news for Indian AI startups than another expensive frontier release: it lowers the cost of building an agentic product on top of someone else's model, which is exactly the layer where most of India's AI startups currently sit.

Markets and AI money

The memory trade keeps compounding in a way that's unusual even by this year's standards.

TrendForce now projects global memory revenue will jump 134% in 2026 to $552 billion, and neither Micron nor SanDisk executives have signaled any near-term easing in hyperscaler orders. It's worth sitting with how unusual that is: memory has historically been the most brutally cyclical corner of semiconductors, prone to gluts as soon as everyone overbuilds at once. The bet embedded in these prices is that this time is structurally different, that AI training and inference simply consume memory at a rate the industry hasn't built capacity for yet. If that's wrong, it's wrong for Micron, SanDisk, and every AI lab depending on their roadmaps at the same time.

That's the day. More tomorrow.