OpenAI's Jalapeno, a 16 GW Battery Grid, and AI in Orbit

The AI buildout has stopped being a story about models and become a story about the physical stack underneath them — the chips that run inference and the electricity that feeds the racks. In a single late-June week, OpenAI and Broadcom revealed a custom inference processor, Qualcomm lined up Meta for its first data-center CPUs, and Tesla, Sunrun and Renew Home pooled more than 16 gigawatts of home batteries into a grid product aimed squarely at data centers. On the fringe, SpaceX put a name to an orbital-compute constellation — the part of this story that deserves the most skepticism.

Custom silicon: the inference land grab

OpenAI's first chip is called Jalapeño, co-developed with Broadcom and pitched as an accelerator built specifically for inference — the compute-heavy job of serving trained models to users, as opposed to training them. According to OpenAI, the part went from initial design to manufacturing tape-out in roughly nine months, a fast cycle the company credits in part to using its own models in the design process. Tom's Hardware describes it as a reticle-sized ASIC. The rollout is staged: small-scale deployment by the end of 2026, larger volumes through 2027, and full-scale operation in the first half of 2028.

Strategically, Jalapeño is OpenAI's move to own the full stack from model to hardware — the same vertical-integration logic Google has followed for years with its TPU line. Owning the inference silicon lets a model maker tune the chip to its own workloads and, more importantly, cut the per-query cost of running models at scale.

Qualcomm is pushing into the same neighborhood from the CPU side. Its Dragonfly C1000 server chip — a chiplet design Qualcomm says carries more than 250 cores and delivers over 2x the performance per watt of existing server CPUs — landed Meta as its first named customer. The catch is timing: the C1000 isn't slated to be commercially available until 2028, so the competitive impact is a few years out.

The common thread is that inference, not training, is now the cost center worth designing custom silicon around — and the hyperscalers and model labs would rather not pay Nvidia's margin on every token served.

The AI Chip Race: Nvidia, Custom Silicon, and the Search for Cheaper Inference

Powering the buildout: a 16-gigawatt virtual power plant

Silicon is only half the constraint; the other half is power, and that is where the most concrete news landed. Sunrun, Renew Home and Tesla announced a deal to aggregate more than 16 gigawatts of distributed energy — hundreds of thousands of home batteries operated by Sunrun and Tesla, plus more than 8 million smart thermostats and devices managed by Renew Home — into what the companies call the largest distributed power plant in the United States.

What makes it structurally different from earlier virtual-power-plant (VPP) programs is the buyer: instead of selling flexibility back to utilities, the consortium sells directly to data-center operators that need fast, dispatchable capacity to cover demand spikes without waiting years for new generation or transmission. In Virginia — the heart of "Data Center Alley" — the partners say more than 300 megawatts is ready for immediate deployment, rising to at least 500 megawatts by 2030. Investors noticed: Sunrun's stock spiked sharply the day the deal was announced.

The appeal is speed. A software-coordinated fleet of existing home batteries can be dispatched like a single generator "without new hardware or land," as the announcement puts it — a meaningful edge when the bottleneck for AI growth is increasingly the grid interconnection queue, not the chips.

The orbital fringe: real hardware, oversized claims

Then there is the part to treat with the most caution. SpaceX has confirmed the name Starmind for a planned constellation of AI satellites that would do computation in orbit on solar power, with Starship envisioned to carry 30 to 50 satellites per launch; Elon Musk has floated a long-term vision of up to a million orbital compute nodes. In parallel, SpaceX flew the first demo of Starfall, a flat, disk-shaped reentry capsule — roughly 3.1 meters wide and about 2,100 kilograms — designed to return up to 1,000 kilograms of cargo from orbit. The first Starfall launched on June 23, 2026 from Cape Canaveral.

The hardware is real; the framing is not. A source headline claiming Starmind would render terrestrial data centers "obsolete" is pure hype and should be read that way. Orbital computing runs into hard physical problems that the press release does not solve: heat rejection (you can't air-cool in a vacuum, so everything radiates slowly), radiation hardening, latency to and from the ground, and the launch cost of putting compute and its power supply into orbit. The distance between naming a constellation and operating one at data-center scale is enormous.

Bottom line

The AI race is now visibly a contest over the stack beneath the models: who designs the inference chips, and who can deliver the megawatts to run them. Jalapeño and the Dragonfly C1000 show custom silicon moving from optional to expected, but both are 2027–2028 stories, not today's. The Tesla–Sunrun–Renew Home VPP is the most immediately useful development, because power — not algorithms — is becoming the gating factor for AI growth. And the orbital constellations are worth watching precisely because they're worth doubting: interesting hardware wrapped in claims that physics has not yet permitted.

Sources and further reading