Good morning. Custom silicon is the through-line today: OpenAI finally has its own chip, Qualcomm just spent $4 billion buying the software stack that might glue alternatives to CUDA together, and Cerebras is being punished by the market for the unglamorous reality of actually building data centers. Around the edges, Anthropic is mad at Alibaba, Google is leaking researchers, and the NSA situation we covered yesterday took another turn.
OpenAI’s Jalapeño arrives, sort of. OpenAI unveiled Jalapeño, its first custom inference ASIC built with Broadcom and fabbed at TSMC, with deployment targeted for end of 2026. Broadcom’s Hock Tan claims roughly 50% cost savings over typical AI GPUs, and OpenAI is emphasizing performance-per-watt and real-time coding workloads, though HN commenters were quick to note the timing lines up suspiciously well with a pre-IPO sales narrative. The other recurring skepticism: with quantization, distillation, and architectural shifts happening this fast, will custom silicon shipping in late 2026 actually earn its ROI before it’s obsolete? One commenter pointed at Taalas, which is going further and burning weights directly into ROM.
Qualcomm buys Modular for $4 billion. Chris Lattner’s Modular is being acquired by Qualcomm, nine months after a $1.6 billion valuation, bringing the MAX inference engine, the Mojo language, and a serious compiler team into a company that’s clearly done being just a phone-chip vendor. Combined with Qualcomm’s earlier Ventana RISC-V purchase, the HN read is that Qualcomm is assembling an end-to-end alternative to the CUDA stack. The irony noted by several commenters: Lattner spent years publicly arguing hardware companies are exactly the wrong people to build AI software stacks, and now he works for one. Others read it as a quiet admission that Modular’s truly-cross-platform pitch was harder than expected.
Cerebras gets punished for renting back its own hardware. Cerebras stock dropped nearly 20% after its first post-IPO earnings, despite Q1 revenue of $193 million (up 94% YoY). The trigger was full-year gross margin guidance of 38–41%, down from 47%, which CEO Andrew Feldman blamed on a temporary arrangement where Cerebras is renting back its own gear from a major customer while it builds out internal data center capacity. Feldman insists the market misread it; the market, evidently, disagrees.
Anthropic accuses Alibaba of distilling Claude. Anthropic says Alibaba ran a large-scale distillation campaign between April and June, generating terabyte-scale outputs through fraudulent reseller networks selling Claude tokens at 70–90% discounts. The HN thread is almost entirely about the hypocrisy, with one commenter summarizing it as “you’re trying to rip off what I’ve already ripped off.” Several pointed out that distillation is, mechanically, very hard to prevent — you can slow it down, but as long as you sell API access, the outputs leave the building.
Gemini 3.5 Flash gets computer use, and a git reset --hard. Google folded computer use into Gemini 3.5 Flash as a native tool, with adversarial training against prompt injection and optional enterprise confirmation gates. The HN reception was rough: one user reported the agent ran git reset --hard instead of committing their changes, another caught Google’s benchmark chart presenting Gemini as a winner despite Opus 4.8 and GPT 5.5 beating it on the same graph, and several complained about over-tuned guardrails refusing mundane questions about SIM transfers and NTFS backups.
Krea 2 ships 12B open weights with an actually-detailed report. Krea released Krea 2, a 12B diffusion transformer text-to-image model under a permissive license, with a technical report covering data curation, RL post-training, and infrastructure in more depth than is usual. Co-founder Diego Rodriguez showed up in the thread to answer questions; early testing suggests the Turbo variant at 8 steps is competitive with Ideogram 4 while being orders of magnitude faster. The main grumble was the choice of Qwen’s VAE.
NSA-Mythos update, and Google’s brain drain continues. Following yesterday’s coverage, the NYT reports the NSA lost access to Mythos amid the unfinalized contract dispute, with some Pentagon officials now pushing the agency toward other models. The HN thread is split between people relieved the NSA doesn’t have it and people noting the “broke into classified systems in hours” claim still smells like marketing. Separately, more senior researchers are leaving Google — Gemini’s Jonas Adler and Alexander Pritzel to Anthropic, following Noam Shazeer to OpenAI and John Jumper to Anthropic. The OpenAI and Anthropic IPOs are doing the recruiting.
That’s the day. Lots of chip news, lots of lawsuits-by-press-release, and a reminder that Cerebras’s “we’re just renting our own hardware back for a bit” is exactly the kind of footnote IPO investors don’t like reading.