Nvidia's numbers fit 15 to 37 GW of 2027 data centers, below the 50 GW forecast

SemiAnalysis forecasts about 50 GW of new data centers in 2027. CTOL's calculation from Nvidia's filing fits 15 to 20 GW with accelerators, or 28 to 37 GW on the widest reading.

SemiAnalysis forecasts about 50 GW of new data center capacity in 2027. A gigawatt (GW) is a billion watts of power. By SemiAnalysis's own words, 50 GW is the world's cumulative stock through 2024 built in a single year.

Capital markets are being asked to finance the buildout. The useful question is how many of those gigawatts can be fitted with accelerators, the chips that run AI, next year. CTOL's answer, built from Nvidia's own numbers, falls well short of 50 GW.

Nvidia's filing gives a way to test the forecast. It says guarantees secure about 4.25 GW at SB Energy's PORTS Pike campus in Ohio. It also says there is an option on about 3.8 GW more.

The filing then gives a revenue figure for the same site. Each generation of Nvidia infrastructure there could represent about 1.5 million GPUs, the graphics processors that do the AI work. That is $150 billion to $200 billion of Nvidia revenue.

Dividing that revenue by the power gives a cost per gigawatt. Against 4.25 GW, it is $35 billion to $47 billion per gigawatt. That also means 2.8 kW of site power behind each GPU.

Against the full 8 GW, the same revenue is $19 billion to $25 billion per gigawatt. Each GPU would then have 5.4 kW of site power behind it.

The filing places that figure after the 3.8 GW option, so it does not say which base applies. CTOL judges the 4.25 GW reading likelier because 2.8 kW of site power per GPU is a plausible all in load. The wider reading needs 5.4 kW, nearly twice as much. CTOL therefore computes both.

Next, CTOL projects Nvidia's 2027 revenue. Nvidia guided its August to October quarter to $108.0 billion. CTOL assumes 8% growth in the next quarter and 5% a quarter through 2027.

CTOL also assumes a 92.5% data center share, the July quarter's ratio. That gives about $488 billion of data center revenue in 2027.

On the likelier reading, that revenue fits 10 to 14 GW of Nvidia systems. On the wider reading, it fits 20 to 26 GW.

Nvidia is not the only chip supplier. If Nvidia supplies 70% of accelerator capacity, a further CTOL assumption, all vendors fit 15 to 20 GW on the first reading. On the second reading they fit 28 to 37 GW.

Filling 50 GW would take $1.75 trillion to $2.35 trillion of Nvidia equivalent systems on the likelier reading. That is 3.6 to 4.8 times the revenue path.

The wider reading would take $0.9 trillion to $1.2 trillion, which is still 1.9 to 2.5 times the path. At a 70% share, chips fall short of 50 GW under both readings.

The share matters. The 4.25 GW reading falls short at any Nvidia share above about 28%, while the wider reading reaches 50 GW at a share of about 52% or lower. At a 70% share, the gap is 30 GW against the top of the likelier range and 13 GW against the top of the wider one.

The buildings themselves are not in doubt. The free section of the SemiAnalysis report counts more than 24 GW in advanced construction at 40 campuses, most of it due in 2027. It calls that a small share of the hundreds of sites at an advanced stage.

CTOL does not dispute the construction count. The disagreement is over what fills the buildings. A hall that is built but empty still has debt, a landlord and a lease clock.

SemiAnalysis names the same chip limit. Its report says memory and logic could become the number one constraint in 2027 and 2028. Its subtitle says the shortage could move to chips. Nvidia's own filing shows supply and capacity purchase commitments of $279 billion, primarily for memory.

CTOL's figures are conservative for a second reason. Older accelerators, CPUs and custom chips from other makers also fill halls, so these are floors for any single vendor. If Nvidia revenue grew 10% a quarter instead, the likelier reading would rise to 17 to 22 GW. The wider one would rise to 32 to 42 GW.

Others have doubted SemiAnalysis's pace as well. Altimeter's Brad Gerstner has argued on the All In podcast that SemiAnalysis's earlier 43 GW figure was too aggressive. He put United States delivery nearer 25 GW, a narrower market than SemiAnalysis covers, according to Yahoo Finance. The same report describes the 50 GW as capacity outside China.

Power is a separate limit, and the two SemiAnalysis reports describe different ground. Its June grid analysis covered the United States. It said headroom, the spare grid capacity, "turns negative by 2027". It also said barely 15 GW of net new firm capacity, meaning power available on demand, is added a year.

That June analysis added that a utility can promise 500 MW for 2027 and later slip it to 2029. It expected behind the meter supply, meaning power made on site, to power well over half of new US data centers from 2028.

The October report is global and more upbeat on generation. It credits fuel cells, truck engines and refurbished turbines for solving the generation problem. It still says grid headroom dips negative at the end of 2027 and that the buildout outside the United States is accelerating.

The equipment record shows how slowly new power arrives. GE Vernova holds 116 GW of gas equipment in backlog and slot reservations worldwide. Yet it shipped only 3 GW in the June quarter and says it is on track for 20 GW of annual output in the third quarter.

Demand is not the doubt. CBRE reported Northern Virginia vacancy of 0.3% in the first quarter, so almost nothing is free to rent. In Dallas Fort Worth a record 716.7 MW was under construction, 88% pre leased.

CBRE also lists grid interconnection timelines in Dallas Fort Worth as not materially improved in the past six to 12 months.

The leases show how much finance leans on the forecast. SemiAnalysis counts more than $1 trillion of leases signed this year, and its subtitle says $1.3 trillion. CTOL uses the lower figure. At $30 billion to $50 billion of contract value per gigawatt, that is at least 20 to 33 GW across 2027 and 2028. The rest of a 50 GW year must come from hyperscaler self build.

Pre leases matter to developers because they cut financing cost. Projects typically risk only 5% to 10% of capital before a tenant signs. But a tenant whose racks arrive late must still be paid from delivery.

Ready for service dates keyed to the tenant's accelerator supply therefore decide who carries the empty hall.

PJM publishes its 2027 long term load forecast in January 2027. Before and after it, each quarterly Nvidia outlook shows whether revenue is compounding near 5% a quarter or nearer 10%. That moves CTOL's likelier reading from 15 to 20 GW toward 17 to 22 GW. On CTOL's 70% share assumption, neither path reaches 50.