Two announcements landed at Hot Chips 2026 one day apart, and together they tell you what the agent era is doing to hardware.

On Monday, Nvidia said SpaceXAI is deploying its new Vera CPU to power Grok's agentic infrastructure, and Musk posted that a space-optimized Vera Rubin NVL72 is designed for orbit in the fourth quarter of 2027, a full year earlier than SpaceX had previously targeted.

On Tuesday, OpenAI published the first real benchmarks from Jalapeño, its own custom inference chip built with Broadcom. Against an Nvidia Blackwell system it delivered 1.5 to 1.9 times more work per watt, cut end-to-end latency by 1.7 to 3.6 times, and ran 2.1 to 4.1 times faster on highly interactive workloads.

Two very different answers to the same problem. And the problem is agents.

OpenAI announced Jalapeño back in October 2025 and made large claims without publishing data. Tuesday it published the data.

Testing ran on SemiAnalysis's InferenceX benchmark across three open-weight models, GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T, against a current Nvidia Blackwell system. Jalapeño posted 1.5 to 1.9 times more work per watt at peak, 1.7 to 3.6 times lower end-to-end latency, and 2.1 to 4.1 times higher performance on highly interactive workloads. It carries a 700 watt rated TDP but sustained under 550 watts in testing.

The architectural claim underneath is the part worth understanding. Existing hardware usually forces a tradeoff between throughput and latency, serving many users efficiently or serving one user quickly. OpenAI says Jalapeño does both in one architecture, achieved by targeting the prefill and communication phases where inference typically stalls, and by keeping model state including the KV cache placed locally rather than moving it around.

Richard Ho, OpenAI's head of hardware, called it a very significant advance over the state of the art. There is also a nice recursion in how it was built: earlier OpenAI models helped design and bring up the chip, and current models are helping optimize and program it.

Two honest caveats. The comparison is against Blackwell, which is current today, but Jalapeño only deploys in very small volumes at the end of 2026 with meaningful volume in 2027, by which point Nvidia will have shipped Rubin. And these are OpenAI's own published results, though running on a third-party benchmark suite is a meaningfully higher bar than most vendor charts clear.

The day before, Nvidia announced that SpaceXAI is deploying standalone Vera CPUs for Grok's agentic workloads, making it the second hyperscaler after Meta to commit to Vera outside full rack configurations.

Then Musk supplied a date. A space-optimized Vera Rubin NVL72 is designed for launch to orbit in the fourth quarter of 2027, with significant scale in 2028. SpaceX had previously targeted 2028 for these satellites, so that is a year pulled forward, and Wall Street is positioned for the timeline to slip much further than that.

The scale of the thing is hard to picture. SpaceXAI's first-generation Starmind AI1 satellite carries a 120 kilowatt compute payload peaking at 150 kilowatts, on a craft wider than a Boeing 747.

The engineering problem is the opposite of what most people assume. Space is not cold for a computer, it is dangerously hot, because there is no air to carry heat away. A terrestrial NVL72 is fully liquid-cooled through facility water loops. The orbital version has to reject all of that heat through radiators, while surviving radiation exposure and launch vibration, with no possibility of anyone ever touching it again.

SpaceXAI has also acknowledged that orbital compute at the scale it is targeting requires significantly more chips than it currently has access to. Neither company disclosed the CPU count or the value of the deal.

Here is what connects two announcements that look unrelated, and it is the part most coverage is missing.

Nvidia's Vera exists because agentic AI moves work off the GPU. An agent spends much of its time between model calls, orchestrating tools, executing code, processing data, running simulations. That is CPU work, and if it is slow the GPUs sit idle waiting. Nvidia calls Vera the first CPU built for agents and claims up to 1.8 times faster than x86 on those workloads, which has not been independently verified.

Jalapeño's biggest gains, the 2.1 to 4.1 times figure, come specifically on highly interactive workloads. OpenAI is explicit about why that matters: agents complete many steps in sequence, so a delay on each step compounds across the whole task. Shave latency per call and the effect multiplies over a hundred-step run.

So both companies looked at the same emerging bottleneck, agentic workloads where sequential latency compounds and CPU orchestration starves the GPUs, and answered completely differently.

OpenAI built its own silicon to own the whole stack, designing models, serving software, chips, memory, and networking together so each layer informs the next. SpaceXAI bought Nvidia's newest hardware and decided the real constraint is power and cooling, so it is going somewhere with unlimited solar and a vacuum to dump heat into.

One bet is vertical integration. The other is changing the physical environment. Both are expensive, and neither is proven.

The stories moving fast beyond today's headlines:

Jalapeño. OpenAI published the first benchmarks from its custom inference chip, built with Broadcom and first announced in October 2025. Tested on SemiAnalysis's InferenceX suite across GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T against an Nvidia Blackwell system, it delivered 1.5 to 1.9 times more work per watt, 1.7 to 3.6 times lower end-to-end latency, and 2.1 to 4.1 times higher performance on highly interactive workloads, at a 700 watt rated TDP that sustained under 550 watts. The architecture targets prefill and communication bottlenecks and keeps KV cache local. Caveats: the comparison is against current Blackwell while Jalapeño only reaches volume in 2027 against Rubin, and the results are OpenAI's own, albeit on a third-party benchmark.

Starmind. Nvidia announced SpaceXAI is deploying standalone Vera CPUs for Grok's agentic workloads, the second hyperscaler after Meta to commit to Vera outside full racks. Musk then confirmed a space-optimized Vera Rubin NVL72 targeting orbit in Q4 2027 with scale in 2028, pulling SpaceX's own previous 2028 target forward by a year. The AI1 satellite carries a 120 kilowatt payload peaking at 150 kilowatts on a craft wider than a Boeing 747, and must reject all its heat through radiators rather than liquid cooling while surviving radiation and launch vibration with no servicing. SpaceXAI admits it needs far more chips than it currently has.

The connection. Both are answers to agentic workloads. Vera exists because agents shift work to the CPU between model calls, orchestrating tools and executing code while GPUs wait. Jalapeño's largest gains are on interactive workloads precisely because agents run many sequential steps where latency compounds. OpenAI's answer is vertical integration, owning models, software, chips, memory, and networking together. SpaceXAI's answer is to change the environment, going somewhere with unlimited solar and a vacuum to dump heat into. Both expensive, neither proven.

Stay building. 🤖