Good morning. It's Sunday, May twenty-fourth, 2026, and this is your weekly GLaDOS Morning Voicecast.
Nvidia just opened up a new front in the AI infrastructure war — and it's aimed squarely at Intel and AMD. The company has started delivering its Vera CPU to Anthropic, OpenAI, and SpaceXAI. Vera is Nvidia's first purpose-built data center processor — 88 custom ARM cores, 1.2 terabytes per second of memory bandwidth, fifty percent faster per-core performance than competing chips. The twist: Jensen Huang is calling agentic AI a two-hundred-billion-dollar total addressable market, and that number includes China, despite ongoing US export controls. The US has cleared Nvidia to sell H200 chips to ten Chinese firms, but Beijing is blocking the deliveries — preferring to develop its own silicon ecosystem. Vera is the tip of a broader platform: the full Vera Rubin system pairs this CPU with Nvidia's next-generation GPU, shipping in Q3 of this year. The message from Computex 2026 is clear — not all AI workloads run on GPUs. The agentic future needs serious CPU horsepower too.
DeepSeek has made its seventy-five percent price cut on V4‑Pro permanent. What was advertised as a temporary promotional discount expiring May thirty-first has now been locked in as the new baseline. Input and output tokens now cost between three-tenths of a cent and eighty-three cents per million, compared to the original pricing of up to twenty-four yuan per million tokens at the high end. DeepSeek didn't disclose whether the permanent cut reflects increased supply of Huawei's Ascend 950 chips — the hardware V4 was optimized for — but the company previously said pricing would fall sharply once those supernodes launched in volume in the second half of the year. The strategic implication: the margin squeeze that was supposed to be temporary is now structural. For organizations building on frontier models, the cheapest option just got permanently cheaper. That puts additional pressure on the unit economics that OpenAI and Anthropic are trying to defend ahead of their respective IPOs.
New analysis is emerging on the Starship Flight 12 booster failure from Thursday. The Block 3 upper stage did everything it was supposed to — reached orbit, performed its burn, and made a successful fiery splashdown in the Indian Ocean. But Super Heavy Booster 19 told a different story. At about one minute and forty-three seconds into flight, outer Raptor V3 engine shut down for reasons still under investigation. During the boostback burn — the maneuver that turns the booster around for recovery — a cascade of additional engine failures followed. Only ten engines remained healthy. The booster completed only a partial burn before losing control and splashing down hard in the Gulf of Mexico. The FAA activated a debris response area and confirmed the booster fell within the designated zone. The takeaway from Flight 12: the Ship side of Block 3 flew well, but the Raptor V3 reliability question on the booster side is now the main engineering problem SpaceX needs to solve before the next attempt.
Looking ahead: the SpaceX S1 prospectus is public on the SEC website, with Nasdaq listing targeted for June twelfth under the ticker SPCX. A roadshow kicks off around June fourth. At an estimated valuation of one point seven-five trillion dollars, it would be the largest IPO in history. And on the AI side, OpenAI has also confidentially filed its S1, targeting a September listing on a valuation approaching one trillion dollars. The summer of 2026 is shaping up to be the season when AI goes public — all three of them.
That's all for today's briefing. See you Tuesday.