In partnership with
⚡ THIS WEEK'S SIGNAL
CoreWeave, the GPU cloud that rents computing power to AI labs, says its near-term capacity is effectively sold out (per coverage of its Q2 call). Its own filings say roughly 41% of the power it has contracted is switched on. Both can be true: selling capacity before the building is energized is the business model. For anyone asking what AI compute really costs, that gap is where the price gets set and the risk gets parked.
━━━━━━━━━━━━━━━━━━━━━━━
📰 MAIN STORY
CoreWeave publishes a ratio most of this industry keeps private. In its Q2 results, the company (Nasdaq: CRWV, public since its March 2025 IPO) reported about 1.5 GW, or 41%, of active power against about 3.7 GW contracted at June 30. Since the 90 GW were announced with only 2 GW running, we now have visibility inside one company's own disclosures. The reason the gap persists is the three-year lead time for transformers: a campus can be leased long before the equipment feeding it can be delivered.
Read what each party is optimizing for. CoreWeave wants power secured ahead of customers, and its 10-K says it expects to commit to long-term power supply contracts at some sites before it has customer contracts for them. Developers want an anchor tenant on a long lease so they can finance construction. The two customers above 10% of Q2 revenue (36% and 26%, unnamed in the filing) want guaranteed capacity, and committed contracts were 98% of revenue. NVIDIA's GPUs fill the racks, and CoreWeave's S-1/A says it typically orders systems when a customer contract is signed. Analysis: that places NVIDIA's order upstream of the energization date. Utilities, in the middle, want no stranded cost if a lease signed in a boardroom never becomes load on a feeder.
Here is the contrast, and it is not a causal one. CoreWeave's operating lease liabilities went from $8.2 billion at the end of 2025 to $16.3 billion at June 30, against a $103.7 billion backlog. A Virginia household using 1,000 kWh a month will see Rider T1 rise 94 cents, per the SCC fact sheet in our issue questioning who pays for new data center grid equipment, with a separate distribution rider proposing roughly $3.46 more. Those are not the same dollars. Analysis: they are two ends of one supply chain, a private lease on one end and a line item nobody reads on the other.
The numbers worth screenshotting, all from filings. Contracted and active power: about 1.3 GW and 360 MW at end-2024 (S-1/A); about 3.1 GW and 850 MW at end-2025 (10-K); about 3.7 GW and 1.5 GW at June 30, 2026 (Q2 release). Remaining performance obligations (contracted revenue not yet delivered) were $103.7 billion, with 41% expected within 24 months, 39% in months 25 to 48, and the rest through month 78 (10-Q, Note 2). First-half capex was $14.1 billion. Recourse and non-recourse debt totaled about $35 billion. Q2 net interest expense was $640 million against $2.575 billion of revenue.
Scenario A: energization closes the gap on schedule. If active power reaches the year-end target management set on its Q2 call, backlog converts to revenue on its disclosed schedule. Analysis: lenders financing the GPUs and developers collecting rent are the investor-side actors that benefit, and utilities see load arrive on the dates in their studies, the version a planner can build a workplan around. Scenario B: it slips. The RPO note itself names delivery delays and availability credits as reasons revenue may not be recognized, while interest keeps accruing. The borrower and its lenders carry that; the operator-side actor is the utility holding a queue position for load that arrives late. For AI buyers, scarcity pricing lasts longer.
One thing to watch: the active-to-contracted ratio in CoreWeave's Q3 10-Q, with 41% at June 30 as the baseline.
━━━━━━━━━━━━━━━━━━━━━━━
🧾 DEAL DECODER
The instrument: CoreWeave's data center lease with its New Jersey joint venture.
Who signs what. CoreWeave as tenant, and a joint venture (JV) formed in June 2025 with a data center developer. It is a real estate lease, not a power purchase agreement. Initial term is 15 years, base rent is a percentage of construction costs incurred, and it commences on completion of construction.
Who fronts the capital. The JV builds with additional debt and equity; CoreWeave's stake rose from 15% to 35%. Rent recovers the build over the term.
Who eats the loss. Before completion, the JV's investors and lenders. CoreWeave discloses maximum exposure of a $44 million carrying value, up to $95 million of guarantee, a $37 million lease prepayment and up to $160 million of construction funding.
Precedent. Rent tied to the landlord's construction bill is renting a house while paying for a slice of the lumber. Expect the next developer to ask for the same.
━━━━━━━━━━━━━━━━━━━━━━━
Discover laptops with built-in intelligence.
Shift your workforce into hyper-productivity mode with Dell Pro laptops powered by Intel® Core™ Ultra with Intel vPro®. Shop on Dell Premier.
━━━━━━━━━━━━━━━━━━━━━━━
⚡ QUICK HITS
Tesla deployed 13.7 GWh of storage in Q3. That is energy stored, not power delivered, and the distinction matters when comparing it with the gigawatts above. The company's Q3 release reports deployments, not buyers, and Q2 was 13.5 GWh per S&P Global. SpaceX bought $295 million of Megapacks in Q2, as we noted in our September 1 issue, so Tesla's October 21 results are where the Musk complex's shared energy appetite next shows up. Storage is the part of the grid story that can be built in months rather than years, which is why it keeps appearing in these filings. [Tesla 8-K, Ex. 99.1] [S&P Global]
Google's 250 MW San Jose resolution was calendared again for October 8. I could not confirm the outcome as of this writing, and the next CPUC (California's utility regulator) voting meeting is October 29. The draft resolution approves PG&E's agreement with modifications requiring additional ratepayer protections, for a load larger than Marin County's average demand, so it is the California counterpart to the Virginia fight. The question to ask of any such agreement is who pays if the customer's load never arrives. [Stoel Rives, Oct. 7] [Draft E-5455]
CAISO's large-load comments close October 15. California's grid operator posted its draft final proposal September 24 and met stakeholders October 1, answering FERC's June 18 show-cause order. Analysis: because the proposal answers a federal order, other grid operators may watch how it lands. For anyone with a California interconnection request in flight, this is the process that decides how large loads get studied, and the comment record is where utilities and developers state their positions in public. [CAISO notice]
Loudoun paused certain data center and substation applications. On September 15 the board voted 7-1-1, with an exception for "grid reliability" substations, and a memorializing resolution is due October 20. The supervisor's release says at least 17 grandfathered cases, over 11 million square feet, are not covered; I could not confirm what happened to the grandfathering question on October 6. Watch the October 20 resolution for the final text, since exceptions in a pause often decide how much it pauses. [Loudoun County release]
🧭 THE EXPOSURE MAP
If you hold equities. Structurally exposed: GPU-cloud business models that sign capacity before it is energized, the lenders financing the hardware, and the developers collecting rent. The events that would confirm or contradict the picture are CoreWeave's Q3 filing and its contracted-versus-active figures. Companies named here are participants, not recommendations. Analysis: the common thread is timing risk, meaning the distance between a signed contract and an energized site.
If you run projects at a utility. Craft knowledge: a tenant's contracted megawatts and a developer's study position are different things. Can your organization tell which one a queue request represents?
If you set or influence policy. Cost-allocation dockets see the utility-facing contract; the lease economics stay private. The next venues are CAISO's October 15 comments and Dominion's direct-assignment filing, due around October 29.
If you pay a utility bill or buy AI compute. Nothing here changes a bill this quarter; rider impacts take quarters to years, and the Virginia figures in our September 29 issue remain proposals and approved riders, not a forecast of your next statement. For AI tools, the mechanism is commercial: capacity pricing flows into vendor price lists on vendor timelines.
━━━━━━━━━━━━━━━━━━━━━━━
🔧 TOOL / RESOURCE OF THE WEEK
CoreWeave's 10-Q for the quarter ended June 30, 2026 (SEC EDGAR): The primary filing behind every number above. Retail investors and credit analysts can read Notes 2 and 3 first, then check the contracted-versus-active ratio themselves. The filing is long, so search for "active power" and "contracted power" to find the definitions first. → sec.gov
━━━━━━━━━━━━━━━━━━━━━━━
💬 CLOSING THOUGHT
I'm cautiously interested. Disclosing contracted and active power side by side is more candor than most of this market offers. For the optimism to hold, energized megawatts have to close on contracted ones on the dates in the filings, and the leases have to stay as legible as this one. Until then, a gigawatt on paper and a gigawatt on a feeder are different products.
Will contracted-versus-energized power ever be reported the same way across the industry?
