บทถอดความฉบับเต็มของ Drax Partnership Plans Phased Data Center — November 3, 2026

DCD reported that Drax, Brookfield and Radiant have signed to develop up to a gigawatt of data center capacity at Drax Power Station, with the first phase targeting up to 100 megawatts and expected to be operational from Q1 2028. The scale is striking, but the first phase still depends on planning and regulatory and network approvals, and on finding a suitable tenant. That puts a very practical question alongside the headline: who pays to make the power and the site ready, and what does the community get? We'll work through what's committed, what remains conditional, and how this deal fits into a day when the biggest numbers are all about infrastructure. Welcome back to Concrete Compute, your daily brief on the AI infrastructure buildout. It's Tuesday, November 3, 2026. Let's get into it. DCD reports that Anthropic now lists aggregate purchase commitments and lease agreements of about $895.4 billion, "substantially all of which is related to compute and infrastructure." This is a single-source account. The company also reported a net loss of $23.29 billion in its third quarter. DCD says compute and infrastructure expenses reached $14.4 billion for the nine months ended Sept. 30, up 329 percent over last year. And the commitments don't all mean the same thing. According to its prospectus, it has committed $121.3bn to lease compute from Google, $120.5bn from SpaceX, $110bn from Amazon, and $44.9bn from Microsoft. The deals with Google, Amazon, and Microsoft are mostly non-cancelable outside of uncured material breaches, and must be paid regardless of actual usage. The SpaceX deal is cancelable with 90 days' notice from next March. One more layer for you: that additional TPU compute expected between 2028 and 2031, could cost $125-150bn, and is not a part of the $895bn calculation. The New York Times reports Anthropic is expected to have more than 3.5GW of compute this year, with a similar amount due in 2027. DCD also reports a $44.6 billion deal with Nscale until 2033 and $41.3 billion with Lambda until 2034; Nscale is required to try to obtain qualifying financing for its Nvidia GPUs, while Lambda will have to secure its facility lease in a timely manner. Big commitments tell us what Anthropic may owe. They don't, by themselves, tell us who ultimately pays for the compute. Staying with the grid that has to feed all that compute, Exelon is talking storage as a bridge while generation catches up in PJM, which is the big regional grid operator across the Mid-Atlantic and parts of the Midwest. Utility Dive - Latest News reports Exelon utilities could spend $3 billion to $7 billion building 1.5 GW to 3.5 GW of regulated energy storage, not including a $1 billion, 500-MW project already proposed for New Jersey. Regulated storage here means batteries owned by the utility and paid for through rates, with regulators deciding what gets built. Atlantic City Electric's 500-MW project, developed with Invenergy, would be the largest battery installation in the PJM Interconnection and is expected to provide more than $700 million in net benefits to customers, he said. Exelon expects a New Jersey decision late in the first quarter next year. The backdrop is tight supply, and Exelon's current plan calls for investing $16.3 billion on transmission through 2029, with room for $12 billion to $17 billion beyond that plan including infrastructure for large loads. Its high-probability large load additions held at 11 GW, with 9 GW in Commonwealth Edison territory and 2 GW in the Mid-Atlantic. On the financial front, Exelon's third-quarter income fell to $729 million, down from $875 million in the same period last year, partly due to higher writeoffs related to a rate case, and higher storm and interest expenses. I think storage helps on peaks and prices, but it doesn't replace the new generation PJM still needs. From the U.S. grid to China, where a long-delayed machine is finally being shown. HPCwire reports the Sugon 8000, also known as Dengfeng, is a 100,000-card supercluster at the National Supercomputing Center in Zhengzhou, hooked into the country's National Supercomputing Internet. Think of a supercluster as tens of thousands of processors wired together to train AI and run science. HPCwire says it uses chips from Hygon, Sugon's scaleFabric interconnect and ParaStor storage, with immersion phase-change liquid cooling for high power-density deployment. Sugon says that Dengfeng will support a range of workloads, from scientific codes requiring full FP64 all the way to running AI models in INT8 mode. Those are just different precision settings, from exact science math to efficient AI math. Sugon says it has already been used for 300 applications. We also don't know if the Sugon 8000 will be listed on the TOP500 list when the latest edition comes out later this month at SC26. For you, the caveat is simple: claimed scale is not a verified benchmark score. And a quick markets note to round the headlines. Proven news breaker @wallstengine posted on X that Howard Marks said if you want to invest in AI, you should probably buy the stocks. A neat lens for today's show, when the equity is chasing gigawatts and the debt is being asked to fund them. That covers the rest of the buildout news. Now, the story I want to spend real time on. Drax is pitching a power-station site in North Yorkshire as a place to develop a very large data center, with Brookfield and Radiant as partners. A data center is a building full of computing equipment; at this scale, power access and approvals are part of the project, not background detail. DCD reports that the partners signed to jointly develop, deliver, and operate up to a gigawatt of capacity at Drax Power Station in Selby, North Yorkshire, UK. The plan is phased. DCD says phase one targets up to 100 megawatts and is expected to be operational from Q1 2028. Phase two would increase capacity to 500 megawatts between 2028 and 2031, and phase three envisions additional capacity