Why AI Project Costs Keep Rising: The Capex Behind Your Compute Bill

Published August 30, 2026By ABD Legacy LLC
ai compute costs AI project costs compute capex AI budgeting

The 30-second answer

NVIDIA's August 17 filing caps its payment obligation at $105 billion under residual value guaranties for OpenAI's ~8 GW Ohio campus — the largest AI computing campus ever constructed. That record AI infrastructure capex is the strongest signal yet that compute cost pressure is structural, not cyclical. The floor under compute and API pricing stays high, so your AI project estimates need a compute-cost floor, not last year's rates.

If you price AI projects off the rates you used in 2025, the August 17 news is your wake-up call. NVIDIA's Form 8-K discloses a $105 billion cumulative cap on its payment obligation under residual value guaranties for OpenAI's initial ~4.25 GW commitment at the PORTS-Pike Technology Campus in Pike County, Ohio [1] — the largest AI computing campus ever constructed, at up to roughly 8 gigawatts [8].

Here is what that capex means for AI compute costs — and why your project estimates need a floor, not a hope.

What the $105B signal actually is

The number is a cap on NVIDIA's payment obligation, not a payment today. NVIDIA entered into multiple residual value guaranties with SB Energy, which will build, own, and operate the campus under a 20-year lease to OpenAI [1][2][3]. NVIDIA pays only on a Trigger Event — OpenAI's insolvency default or lease-payment failure — and OpenAI reimburses and indemnifies NVIDIA for amounts actually paid [1]. Huang's on-record answer to the circular-financing question: "Is this circular financing? No. OpenAI will pay the lease." [8][9]

What matters for your budget is not the guarantee mechanics — it is the demand signal underneath them. OpenAI, the largest buyer of frontier compute, is committing to a 20-year lease at this scale [3]. That is not speculative demand; it is contracted demand, and it competes with your project for the same compute.

Why the floor under compute pricing stays high

AI compute costs are driven by supply, and the supply side of this deal is enormous — and expensive to build. The campus comes with at least 10 GW of new energy generation and at least $4.2 billion in regional grid infrastructure with AEP Ohio [2][4]. The first 800 MW is expected in 2028 on existing AEP infrastructure, with the full campus targeting the end of the decade [3][8]. Every gigawatt of that buildout carries financing, construction, and power costs that eventually show up in per-token and API pricing.

The news coverage framed the number correctly — CNBC reported the $105 billion financing on August 17 [4], Bloomberg detailed the 20-year lease and ~8 GW capacity [7], Fortune noted the deal came in $145 billion lower than the originally reported figure [6], and TechTimes corrected the July $250B reports against the filing [8]. What the coverage did not answer is the agency-buyer question: what should I charge, and what should my client pay, when the capex behind the compute bill is this large? That is the gap this page fills.

What this means for your estimates

If your pricing math assumes compute keeps getting cheaper, this deal is evidence you are wrong on the input side. Model-specific price cuts still happen — but they are events, not a trend you can build a quote on. The structural signal is the opposite: at this scale, the cost floor under compute and API pricing stays high.

The takeaway: AI project costs keep rising because the capex behind them keeps rising. NVIDIA just committed the largest AI infrastructure campus ever built to a 20-year lease with the largest AI buyer [1][2][3][8]. Agencies that price with a compute-cost floor protect margin; agencies that price off stale rates eat the difference when the invoice arrives. For the full capex story and its budget-defense playbook, see the sister guide on AI data center investment and client AI budgets.

Get a compute-aware estimate for your next AI project

Open the AI Agency Cost Calculator →

Related reading: AI model cost per task 2026 · hidden costs of AI automation · how to budget for AI automation

Frequently asked questions

Are AI compute costs going up in 2026?

The structural signal says the floor stays high. NVIDIA's August 17 filing caps its payment obligation at $105 billion under residual value guaranties for OpenAI's roughly 8 gigawatt Ohio campus — what TechTimes calls the largest AI computing campus ever constructed — with OpenAI paying the lease under a 20-year term. At this scale, the cost of building and financing AI infrastructure keeps a floor under compute and API pricing. Individual model prices can still fall for specific tiers, but the input-cost pressure behind your AI project costs is structural, not cyclical.

What does the $105B Ohio deal mean for my AI project budget?

It means your estimates should carry a compute-cost floor, not last year's rates. Record AI infrastructure capex — 35,000 construction jobs, at least 10 GW of new generation, $4.2 billion in grid infrastructure — signals sustained demand for the same compute you buy. If you price per-token and API work off stale rates, mid-project cost drift eats your margin. Build in a compute-cost floor and re-check the rate assumptions in your project estimate before quoting.

Is NVIDIA really spending $105 billion?

No. The $105 billion is a cumulative cap on NVIDIA's aggregate payment obligation under residual value guaranties with SB Energy, the developer that builds, owns, and operates the campus under a 20-year lease to OpenAI. NVIDIA only pays if a Trigger Event occurs — OpenAI's insolvency default or lease-payment failure — and OpenAI reimburses and indemnifies NVIDIA for any amounts actually paid. The separate equity piece is $1.5 billion in SB Energy. The guarantee is the demand signal; the lease payments are OpenAI's.

How should I price AI projects when compute costs are rising?

Price off current per-token rates with a compute-cost floor, add a buffer for mid-project drift, and re-check your assumptions quarterly. The Ohio deal is a reason to treat compute as a structural cost line, not a discountable line item. Use a cost-per-task benchmark for the models you actually deploy, and add a capacity/availability buffer to your three-year TCO where supply is tight. Label any buffer as an estimate — the goal is margin protection, not a fake price list.

Will AI model prices keep falling if data centers get cheaper to run?

Not necessarily. More efficient models and competition push some tiers down, but the infrastructure being built now — roughly 8 GW of capacity, at least 10 GW of new generation, a 20-year lease with OpenAI as payer — locks in demand for the compute that underlies API pricing. Efficiency gains can lower per-token cost for specific models, but the floor under the input cost stays high. Treat published price cuts as model-specific events, not a trend you can build a quote on.

Related reading

Sources

Accuracy note: all figures are from the verified fact sheet for this story (kanban t_16ee40e6), traced to NVIDIA's Form 8-K and Exhibit 99.1 (SEC EDGAR), OpenAI's announcement, CNBC, Bloomberg via Yahoo Finance, Fortune, TechTimes, and Axios, verified Aug 2026. The $105B is a cumulative cap on NVIDIA's aggregate payment obligation for its initial ~4.25-GW commitment, payable only on a Trigger Event and reimbursed/indemnified by OpenAI; it is not a $105B investment, not a loan, and not money NVIDIA pays now. The separate equity piece is $1.5B in SB Energy. Press-reported figures not in the filing — including the $500B total-project-cost estimate and the $250B backstop discussion figure — are labeled as reported, not repeated as facts. The NYT article URL is cited from indexed-search corroboration; the page sits behind a paywall/bot wall (DataDome) and its body could not be fetched at verification time — no load-bearing claim in this article relies on it. Pricing guidance (compute-cost floor, drift buffer, quarterly re-check) is operational guidance, not a published price list.