Luxor Energy Cuts Nvidia B200 Power to 25% in 0.5 Seconds, Probing Bitcoin (BTC) Miners' Grid Role

A 0.5-Second Throttle to a Quarter of Power

Texas data center economics just got a stress test with a stopwatch. Luxor Energy and Bentaus ran an experiment on a single Nvidia B200 chip showing that inference workloads can be throttled to roughly 25% of normal power draw in just 0.5 seconds. Throughput dipped while the restriction was active, but in-flight jobs survived intact and the chip returned to its original speed once the limit lifted — no lost work, no restarts. The motivation lies in a quirk of the Texas power market. Each month from June through September, the grid operator ERCOT records the single 15-minute interval of highest system demand — the so-called 4CP window — and those four snapshots set a large customer's transmission charges for the following year. Under 2026 regulatory figures, ERCOT's transmission costs total about $6 billion; average 4CP demand ran at 80,859.8 megawatts, translating to roughly $74.89 per kilowatt annually. For a 100-megawatt site, that is about $7.49 million a year — a line item as decisive for big power consumers as Brent crude pricing is for oil producers. Notably, Luxor's reduction was not ordered by the grid; the company predicted the peak and curtailed on its own initiative. A 0.5-second response is overkill for a 15-minute window — its real value lies in faster balancing and ancillary markets. Bitcoin mining, the energy-intensive process that secures the network through its proof-of-work consensus mechanism, has long supplied exactly this kind of price-responsive, switch-off-able demand, and the open question the experiment poses is whether AI workloads can inherit it. The urgency is real: ERCOT logged a provisional demand record of 91,089 megawatts on July 22, and Texas Governor Greg Abbott said in August that 474 gigawatts of new connection requests are under review — roughly 90% of them data centers. Only about 205 gigawatts cleared preliminary screening in a July 28 provisional review, and new transmission lines take four to eight years to build even in advanced economies.

NVIDIA Commits $3.5 Billion to MediaTek

NVIDIA has put $3.5 billion behind a decade-long chip alliance with MediaTek, purchasing convertible bonds issued by the Taiwanese semiconductor firm. The companies' chief executives — Jensen Huang and Rick Tsai — laid out the plan in a Bloomberg interview broadcast this week, describing an effort to pool AI supply-chain resources by combining NVIDIA's NVLink interconnect with MediaTek's strengths in SoC (system-on-chip) design and custom XPU accelerators. The first flagship product of the pairing already exists: DGX Spark, a desktop AI supercomputer rated at 1 petaflop that runs advanced generative models locally rather than in the cloud. The two firms are also working with Microsoft on a new processor intended to redefine Windows PC performance for an era of AI agents. On the data center side, MediaTek's custom silicon will connect to NVIDIA's AI factory stack, including Spectrum-X networking, shortening enterprise customers' path to deploying purpose-built chips. Huang moved to shut down circular-financing concerns raised about the deal's size, noting that MediaTek is highly profitable and the world's largest SoC vendor by shipment volume, with each company running independent businesses and clearly defined cooperation scopes. NVIDIA's finance chief separately characterized the transaction as a long-term, equity-style strategic investment in a strategic partner. MediaTek's Tsai said the proceeds will fund research and supply-chain consolidation: expanded ASIC design capability, interconnect IP development and CoWoS advanced packaging, with NVIDIA supplying the custom base-layer technology for next-generation high-bandwidth memory (HBM) so the two companies can share suppliers and push specification standardization. The technical centerpiece is NVLink Fusion, a chip-to-chip interface that lets MediaTek SoCs and custom accelerators link directly to NVIDIA GPUs, sharing memory and network resources without customers building their own switching fabric — cutting development thresholds and time-to-market for hybrid AI factories. The partnership also extends into consumer hardware through the RTX Spark AI PC line. Readers tracking the market in real time can follow live spot and futures prices on Bitget.

Both stories describe the same bottleneck from two ends: not enough power, and silicon that burns plenty of it. The companies' joint announcement confirms the deal's concrete terms — $3.5 billion via convertible bonds, a ten-year engineering roadmap, NVLink Fusion integration and the 1-petaflop DGX Spark — but discloses neither the notes' conversion terms nor any revenue targets, a gap investors should resist filling by inference. For miners, our reading is blunt: the Texas test is the first credible demonstration that AI inference can behave like the curtailable load Bitcoin miners have sold to grids for years. Decentralized projects such as Grass, which monetizes spare bandwidth, and private inference networks like Venice are pushing the same flexibility to the edge, meaning the contest for grid-relevant demand response is only beginning.