CoreWeave’s $40M Per Megawatt Is Not What the Fleet…

The $40 million per megawatt that CoreWeave (NASDAQ: CRWV) disclosed on 17 September is not the new price of AI compute. It is the rate on contracts lasting three to six months, and on management’s own year-end guidance the whole fleet will earn roughly $10 million per active megawatt, about a quarter of the headline. The $4.2 billion convertible that pays for more of that fleet is also dearer than its coupon suggests: counting the $566.2 million capped call, the cash cost works out near 5.7% a year if the stock never reaches $97.85, roughly double the 2.875% printed on the notes. Put the two numbers together and the investment case gets narrower and more testable. It rests on how many megawatts CoreWeave can sell at short-contract prices before the rest of its book reprices, and the company has not disclosed that share.

Here is the arithmetic nobody has laid out per megawatt. In the second quarter CoreWeave booked $2.575 billion of revenue on 1.5 gigawatts of active power, which annualises to about $6.9 million per megawatt. Against each of those megawatts it carried roughly $3.7 million a year of depreciation and $1.7 million of interest, before paying for electricity, data centre leases or staff. A megawatt sold at $40 million covers that stack several times over; a megawatt sold at the fleet average barely covers the machinery and the debt. That is why the per-megawatt figure matters more than the convert, and why the sell-side argument over CoreWeave is really an argument about mix.

What $40 Million Per Megawatt Actually Measures

The figure comes from a single sentence in the prospectus supplement CoreWeave filed alongside its new at-the-market share programme. Since 30 June, the company said, it has “signed short-dated customer contracts at pricing of approximately $40.0 million per megawatt, calculated as annualized revenue divided by power required to service the related clusters, with terms of approximately three to six months.” The accompanying investor slide labels it a “recent pricing range for signed short-dated contracts (3-6 months) in Q3”.

Three details in that definition change how it should be read.

The denominator is the power a cluster draws, not the power a building is connected to. The numerator is annualised, so a four-month contract at that rate produces around $13 million of actual revenue per megawatt, not $40 million. And the population is short-dated deals only, which is a small part of a book where committed contracts, typically around five years long, produced 98% of second-quarter revenue.

Converted into the unit a power trader would recognise, $40 million per megawatt-year is about $4.57 per kilowatt-hour of cluster capacity. The average US industrial customer paid 9.17 cents per kilowatt-hour in June 2026, according to the US Energy Information Administration. CoreWeave is therefore charging roughly 50 times the price of the electricity for the privilege of having that electricity flow through Nvidia silicon for a few months. The gap is the price of scarcity: GPUs that are installed, networked and available now rather than in the 12 to 18 months CoreWeave’s own presentation says a new data centre takes to build.

Pricing on the rest of the book has risen too, just not to that level. The same presentation reports a price increase of about 25% across SKUs in July 2026, expected contribution margins 5 to 10 points higher on second-quarter contracts than in recent quarters, and customer prepayments on about 70% of deals signed in the second quarter. Those are real improvements. They are also improvements measured against a base that, per megawatt, sits far below the short-contract rate.

Chief Financial Officer Nitin Agrawal described the timing problem on the second-quarter earnings call: “A typical 5-year contract carries strong and still-expanding unit economics across its term. But those economics do not arrive evenly.”

Who Has Responded, and What Has Gone Unsaid

Management has been explicit that short contracts are a deliberate profit lever, not an accident of demand. On the same call, Chief Executive Michael Intrator said the delayed-draw term loan closed in August, the first CoreWeave facility to finance shorter customer contracts, lets the company “populate the curve” of contract tenors. Agrawal added that it lets CoreWeave “grow our exposure to shorter dated contracts that typically come at a higher ASP and margins.” Intrator was blunter: “We want to sell our compute on long-term contracts. We also want to sell it on shorter-term contracts to extract additional margin. And we have been really, really aggressive about doing that.”

What management has not said is how many megawatts are on those terms. CoreWeave gives the price of short-dated deals, but no volume, no revenue share and no count of contracts. Its 10-Q splits revenue only into committed and on-demand, and a three-month take-or-pay contract counts as committed. Until the company discloses the mix, the $40 million figure describes the top of the curve without saying how much of the fleet sits there.

The sell side has pressed on exactly this point. Michael Turrin, analyst at Wells Fargo Securities, told management on the August call that despite roughly 500 MW of new active power, “the revenue, if we’re looking at the sequential adds is fairly consistent with last quarter.” Agrawal’s answer was that 300 MW of the additions arrived in June alone, so the revenue would show up in the third and fourth quarters. That is a fair explanation, and it cuts both ways: it means the second-quarter per-megawatt figure is flattered downward, but it also means the fleet average has not yet been tested at full utilisation.

Rothschild & Co Redburn has taken the other side, initiating CoreWeave at Sell with a $54 target on 21 September, an analyst call FinanceFeeds covered in its piece on Redburn’s Sell against the convert strike. Redburn’s first pressure point, falling GPU rental prices, is a direct challenge to the idea that short contracts can hold near $40 million per megawatt.

Then there are the banks. A group of option counterparties including Barclays, Citibank, Deutsche Bank, Goldman Sachs, HSBC and Wells Fargo sold CoreWeave the capped calls. They are paid to take the other side of the dilution protection and, as CoreWeave’s 10-Q warns about its earlier deals, typically hedge through derivatives and trades in the stock itself. None has commented publicly, which is normal. The point is that the most sophisticated counterparties in the deal priced CoreWeave’s volatility, not its per-megawatt economics.

The Per-Megawatt Table: Five Ways to Price the Same Fleet

Line up every per-megawatt figure CoreWeave has disclosed, directly or through arithmetic on its own filings, and the $40 million number sits at one end of a very wide range.

Measure	Per megawatt	Basis and source
Short-dated contracts (3–6 months)	~$40.0M a year	Company figure, prospectus supplement, 17 Sep
New commitments, 30 Jun to 11 Aug	~$50M total contract value	More than $25B over ~500 MW of added contracted power; not a like-for-like match
Backlog at 30 June	~$28.2M total contract value	$104.2B over ~3.7 GW contracted
Guided year-end run rate	~$10.0–10.5M a year	$18.5–19.5B exit run-rate revenue over 1.85 GW active power
Q2 revenue, annualised	~$6.9M a year	$2.575B x4 over 1.5 GW active power
Installed equipment, gross	~$26.5M	$33.8B technology plus $6.0B data centre equipment over 1.5 GW (10-Q)

Two readings follow. First, the new commitments are richer than the old book. Roughly $50 million of contract value per added megawatt against $28 million for the book at 30 June is consistent with the 25% price rise and wider margins CoreWeave reported, though the company does not tie the $25 billion to specific megawatts, so the match is loose.

Second, and more important for anyone valuing the backlog, short-dated contracts barely register in it. A five-year contract at $10 million a year adds $50 million of backlog per megawatt. A four-month contract at $40 million a year adds about $13 million. The most profitable deals CoreWeave signs are the ones that do least for its headline backlog figure, which suggests most of the $25 billion of new commitments came from longer-dated contracts rather than short-term buyers. Agrawal also said that more than 50% of the $104.2 billion backlog was attached to contracts where delivery had started, rising to more than two-thirds by year end. Roughly half the backlog, in other words, was still waiting on capacity at 30 June.

The investment side of the table is the sobering one. Gross equipment of about $26.5 million per active megawatt means a megawatt rented at $40 million a year repays its hardware in well under a year of revenue. At the second-quarter fleet average, the same megawatt needs close to four years of revenue just to match its equipment cost, before interest. Much of the equity story depends on which of those two paybacks describes the marginal megawatt. FinanceFeeds’ coverage of CoinShares’ per-megawatt comparison of AI compute and bitcoin mining shows how quickly that metric has become the sector’s common yardstick.

The Financing Tension: What the Capped Call Really Costs

The convert closed on 22 September at $4.2 billion, including the full $500 million option, after launching at $3.0 billion. Its headline terms are cheap: a 2.875% coupon and conversion at about $97.85, a 22.5% premium. The capped call changes the maths. CoreWeave spent $566.2 million, 13.5% of principal, buying back the dilution between $97.85 and $199.70 a share.

Treat that premium as a cost of the money and the picture shifts. CoreWeave received $4,137.0 million after the initial purchasers’ discount and paid $566.2 million straight back out, leaving about $3,571 million, or 85% of face value. Against that cash it owes $120.75 million of coupons a year and $4.2 billion at maturity on 1 April 2033. A FinanceFeeds calculation of the internal rate of return puts the cash cost at about 5.7% a year if the notes are never converted, before offering expenses. Without the capped call, the same calculation gives about 3.1%.

That is still far below the 9.625% on CoreWeave’s 2032 senior notes. The comparison with the April convert shows the same protection getting dearer.

Term	1.75% notes due 2032 (April)	2.875% notes due 2033 (September)
Principal	$4.0B	$4.2B
Conversion price	~$119.60	~$97.85
Capped call cap price	$230.00	$199.70
Capped call cost	$492M (12.3% of principal)	$566.2M (13.5% of principal)

In five months the coupon rose 112.5 basis points, the conversion price fell 18%, the cap fell 13% and the hedge cost more per dollar raised. None of that is a crisis. It is a steady repricing of CoreWeave’s cheapest capital.

The dilution maths is where the capped call earns its cost. At the initial rate, the notes convert into 42.9 million shares, 7.8% of the 551.5 million Class A and Class B shares outstanding on 31 July. The 52.6 million figure circulating on social media is the make-whole maximum, not the base case. Assuming CoreWeave settles principal in cash, the capped call means no net new shares until the stock passes $199.70. At $250, net issuance would be about 8.6 million shares, against about 26.1 million without the hedge.

The capped call protects shareholders only if the stock rises. If it stays below $97.85, the $566.2 million is simply spent. Meanwhile the separate programme to sell up to 35 million new shares cannot be used until 30 days after the convert’s purchase agreement, which points to mid-October at the earliest.

What Happens Next

Third-quarter revenue per megawatt will rise, but nowhere near $40 million. CoreWeave guided third-quarter revenue of $3.45 billion to $3.6 billion. If active power averages somewhere around 1.65 GW as it moves towards the year-end target of more than 1.85 GW, the midpoint annualises to about $8.5 million per megawatt. The June power catches up and the July price increase starts to show, so the figure improves. It still leaves a gap of roughly four to five times against the short-contract rate, which is the clearest sign that short deals remain a small slice of the book. I expect management to keep quoting the $40 million figure and to keep declining to size it.

The CME compute curve will price the short end in public. CME Group and Silicon Data plan to list H100 and B200 rental futures on 5 October, pending regulatory review. Those contracts track hourly neocloud rental indices, which is closer to CoreWeave’s three-to-six-month deals than to its five-year book. If the forward curve slopes steeply down, the market will be saying the $40 million rate is a peak, not a plateau, and CoreWeave’s short-contract strategy becomes a timing trade. FinanceFeeds explained how the contracts work in its preview of CME’s GPU rental futures.

Equity issuance will be the tell on financing cost. Once the lock-up ends, every share sold at around $80 is equity raised at less than half the $199.70 cap the company just paid $566.2 million to defend. Heavy use of the at-the-market programme in the fourth quarter would signal that management prefers dilution now to more 9%-plus unsecured debt, and would confirm that the 5.7% convert was the cheapest money on offer. Light use would suggest the short-contract cash flow is arriving as promised. The broader neocloud selloff, which FinanceFeeds tracked when CoreWeave and Nebius fell about 20% in a month, shows how little patience the market has for either signal to take time.

FAQ

What does CoreWeave’s $40 million per megawatt figure mean?

It is the pricing on short-dated customer contracts, about three to six months long, signed since 30 June 2026. CoreWeave calculates it as annualised revenue divided by the power needed to run the related GPU clusters. It describes a subset of new deals, not the average revenue the whole CoreWeave fleet earns per megawatt, which is far lower.

How much does CoreWeave earn per megawatt across its fleet?

Second-quarter 2026 revenue of $2.575 billion on 1.5 GW of active power annualises to about $6.9 million per megawatt. Management’s guidance of $18.5 billion to $19.5 billion of exit run-rate revenue on more than 1.85 GW implies roughly $10 million to $10.5 million by the end of 2026. Both are FinanceFeeds calculations from company figures.

What is the capped call on CoreWeave’s convertible notes?

It is a set of option contracts CoreWeave bought from a group of banks for about $566.2 million. They offset dilution from the $4.2 billion of 2.875% notes due 2033 if the share price ends up between the $97.85 conversion price and a $199.70 cap. Above the cap, dilution resumes. Below the conversion price, the capped call expires with no value.

How many shares could the new CoreWeave convertible create?

At the initial conversion rate of 10.2194 shares per $1,000, the $4.2 billion of notes covers about 42.9 million shares. The maximum under make-whole provisions is 52,578,540 shares. With principal settled in cash and the capped call in place, net new shares would only be issued if the stock traded above $199.70.

Does the $25 billion of new commitments appear in CoreWeave’s backlog?

No. The $104.2 billion revenue backlog reported for 30 June excludes more than $25 billion of net new customer commitments added between 30 June and 11 August. Those should appear in the third-quarter backlog figure, due with CoreWeave’s next quarterly results, alongside any new contracts signed since.

Disclaimer: This article is analysis based on public company filings and statements. It is not investment advice, and nothing in it is a recommendation to buy or sell any security. Per-megawatt and cost-of-capital figures are FinanceFeeds calculations from disclosed data and rest on the assumptions stated.