Ten years. That's how long Amazon, Google, Meta and others reportedly told White House officials it could take them to stop relying on Chinese optical transceivers. And the administration's answer, per Reuters, is basically: start counting. Welcome to The Data Center Daily. It's Thursday, November 19th. I'm Sarah, here with Matt. Four items today: the transmission owners' own show-cause filings and the long-term contracts they want data centers to sign, Goldman Sachs nearly doubling its AI project finance forecast, that transceiver import ban moving forward, and OpenAI's Georgia campus clearing a county vote over a room full of objections.
Nick Zenkin, writing for Latitude Media:
And while each takes a different approach to answering that question, together they are asking FERC to endorse an approach where data centers sign a long-term contract with the utility that owns the relevant lines before the upgrades it needs can get built.
We've spent the week on what the grid operators filed under FERC's show-cause orders. Five of the six had answered by Monday, and Southwest Power Pool owes its filing tomorrow. Zenkin's point is that the utilities that own the wires filed their own proposals too, grouped by region, and those answer a different question. Not how fast a load connects, but who pays for the upgrades, and for how long. The terms. PJM's transmission owners want a cost recovery agreement signed directly with the data center: pay at least 80 percent of expected transmission charges for a minimum of ten years, with a financial guarantee that starts at 25 percent of what's owed and rises to the full amount by the time service begins. Leave early and you owe the rest. California and New York would run 15 years. MISO's is the three-way contract we walked through yesterday, running up to 20. Latitude Intelligence lined those up against the state tariffs the same utilities already run. PJM utilities with a state transmission commitment average a 12-year minimum, so the federal floor of ten is actually shorter. Dominion's Virginia tariff, which takes effect in January, sets 85 percent of contracted transmission demand over 14 years. AEP Ohio is 85 percent over 12. In MISO, state terms average 15 years against the proposed 20, and California and New York tariffs set no minimum term at all. Here's where it bites. In retail choice states like New Jersey and Pennsylvania, the customer buys transmission through a power supplier, so a utility minimum bill generally doesn't apply. The federal agreement becomes the data center's only long-term commitment on transmission upgrades. And PG&E reports a 12.9 gigawatt data center pipeline with no minimum term and no minimum bill at retail. So in Northern California it's 15 years on the federal side and nothing on local wires or power plants. At least until California's September law requiring large load tariffs by 2028 takes hold. None of this is final. Comments on several filings are due December 7th, and FERC rejected most of PJM's last large load plan in October, partly for reaching into retail territory. The PJM owners also asked that existing utility contracts with comparable protection stand in for the new one. So a Virginia data center keeps the deal it has, and a Pennsylvania one signs the federal version. Patchwork. Also filed Monday, and easy to miss: NERC's petition for three new reliability standards for what it calls computational loads, per the American Public Power Association. CLO-001, 002 and 003: interconnection studies and modeling, operational data and a duty to follow operating instructions, and protection coordination with disturbance monitoring. It answers FERC's July 16th order, and NERC wants expedited action. Meaning the data center itself becomes an entity that has to comply with NERC standards. For operators, that's a bigger cultural shift than any 80 percent minimum bill.
Dataconomy, reporting Wednesday on a forecast from Goldman Sachs' credit team:
A large portion of the projected $320 billion in chip financing may lack traditional corporate guarantees, Goldman said. In such arrangements, lenders rely on the customer contract and the customer’s creditworthiness rather than a parent-company commitment to repay the debt.
The headline number: Goldman now sees 545 billion dollars of AI project finance in 2027, up from a 300 billion estimate. About 225 billion of that is shells, meaning buildings, power connections and physical infrastructure, and 320 billion is chips. Chip financing has reached 48 billion so far this year, so Goldman expects almost sevenfold growth in a single year. The structure matters more than the size. A special purpose vehicle owns the chips, a customer signs a lease or service contract for compute, generally five years or less, and lenders get repaid out of that cash flow. A lot of it without a parent guarantee. So what the lender is really underwriting is the tenant. Goldman also raised its 2027 hyperscaler capex estimate to between 1.2 and 1.4 trillion dollars. At the top of that range, annualized spending on AI hardware and intellectual property would equal nearly 4 percent of US GDP. Dataconomy cites a reported 40 billion dollar GPU financing package involving SpaceX, backed by Goldman and Apollo. Hold that up against yesterday. The Fed counted more than 3 trillion dollars of commitments sitting off big tech's balance sheets, and HSBC sold Project Jupiter construction debt below 90 cents. Goldman's forecast says the slice without a corporate backstop is about to get a lot bigger. For the customer-contract side, a live example from Datacenter Intelligence on Tuesday. OpenAI signed a six-year agreement with TensorWave for 457 megawatts of dedicated capacity across several US facilities: AMD Instinct MI455X GPUs in Helios racks on Ethernet, first capacity in the first half of 2027 and full rollout in early 2028. The publication calls TensorWave the first neocloud publicly confirmed to support the OpenAI and AMD collaboration. Six years, single tenant, hundreds of megawatts. That's precisely the contract an SPV lender wants to see, and precisely the one that has to hold up. Goldman lists the same constraints I would: power supply, regulatory opposition, and whether end users generate enough revenue to meet those contracts.
Alexandra Alper and David Shepardson, reporting for Reuters:
US firms could need up to a decade to comply with any such mandate, the executives told the officials, according to two people familiar with the matter.
The mandate is a Federal Communications Commission effort, first reported by Reuters in August, to ban US imports of new models of Chinese optical transceivers, the parts that move data over fiber inside the data center. Amazon, Google, Meta, CoreWeave and Cisco joined Microsoft in a White House meeting earlier this year to push back. Reuters reported Tuesday, citing three more people, that the measure is moving forward anyway, though it could still be shelved if sentiment in the administration changes. Why a decade? Counterpoint Research puts China at two-thirds of the global transceiver market. Eoptolink and Zhongji Innolight dominate supply and count top US hyperscalers as customers, and Innolight is on the Pentagon's list of Chinese military-backed companies. Coherent and Lumentum compete, but per a Foundation for American Innovation report, they lack the scale to replace those vendors. The companies also raised the prospect of retaliation, specifically a Chinese export ban on indium phosphide, a critical mineral US firms need to make their own transceivers. Coherent rose 2.2 percent and Lumentum 1.5 on the report. China's embassy said Beijing firmly opposes the US overstretching the concept of national security. And the momentum runs straight through our Nvidia thread. Reuters credits it partly to what it calls the revelations about Jensen Huang's affair with an alleged Chinese spy, and to chip smuggling cases it says the Commerce Department left unaddressed. If your network build has a 2027 date on it, ask your optics vendor this week where the parts come from.
From the Effingham Herald in Springfield, Georgia:
Effingham County commissioners unanimously approved OpenAI's roughly 1,400-acre Project Camellia data center site plan Tuesday night, clearing a major hurdle for the development despite strong opposition from some residents at the meeting.
That vote came Tuesday, November 17th. The Atlanta Journal-Constitution puts the campus at 20 billion dollars and 4 million square feet. The plan now shows six buildings instead of four. Project engineer Aakash Verma said that split doesn't add power or water use, and that separate buildings make it easier to curtail load if they ever need to. Curtailing load, offered up unprompted at a county meeting. That's the grid conversation reaching the zoning room. Commission chairman Damon Rahn, a Southern Company employee, recused himself, since Southern owns Georgia Power, which will supply the site. Verma also confirmed four diesel generators per building, and said they protect life safety systems and help restart equipment after an outage, not run compute. The vote was unanimous among the five participating commissioners, over boos and signs reading Sink OpenAI, Save our water. A moratorium on future data centers was tabled until January on a 4 to 1 vote, and as drafted it would exclude Project Camellia anyway. Commissioners also held a first reading to make data centers a conditional use, so future projects would get the public zoning review this one didn't. And a site plan isn't a building permit. OpenAI filed a Section 404 application on October 5th to fill about 72 acres of forested wetlands, roughly 50 acres more than the whole industrial park was originally permitted for. The Army Corps comment period now runs to December 3rd. OpenAI says it will put more than 20 million dollars into wetland mitigation. For context, this is the state where, as we covered on the 14th, Governor-elect Keisha Lance Bottoms has promised a temporary ban on approvals and the PSC flipped blue. A county site plan in hand counts for something in that climate.
If the TensorWave deal made you wonder what all that compute is for, our sibling show covers that end. Lantern AI Daily Briefing brings the top AI news for engineers, founders and investors every weekday, separating real capabilities from demo hype, explained fast. That's where you find out whether the models can pay for six-year contracts.
Tomorrow is Friday, and the last of the six show-cause filings, SPP's, is due. That's our first stop if it lands. Bring your redlines on that PJM cost recovery agreement. You've been listening to The Data Center Daily, part of the Lantern Podcast lineup. From Sarah and Matt, thanks for the time.