Re-run your 2027 GPU budget: five checks when compute appreciates
When B200 residual values and rental rates rise, re-run depreciation, $/token, break-even, inflation, and renewal leverage before signing capacity.

You are deciding whether to sign GPU capacity now or re-run the budget after the market moves. Start with the arithmetic: B200 GPU-hour rate × GPU count × hours × days is the monthly compute line before power, network, and labor. Rising residual values and rental rates make that line unstable. A budget that ignores the move will be revised after the fact.
Compute is appreciating, not depreciating
Depreciation schedules assume GPUs lose value. B200 resale clears above book value, so the asset carries an exit price. That changes the budget because the exit price affects both the buy case and the rent case. The exit price is a market number, not a schedule entry.
Residual value and rental price are both moving up. Early signers may lock a lower rate; late signers may face a higher rate and a larger upfront payment. The five checks below force that comparison before signature, not after the contract starts.
Run the five checks before signature
Each item produces a number you can defend in a budget review and a decision rule for the next contract.
- Residual-value depreciation: B200 residual value stood at 1.58 times launch cost, a 58% premium, as of September 16, 2026. Older NVIDIA A100 and H100 GPUs have resale values above the book values implied by 3-year or 5-year straight-line depreciation. Check depreciation against resale, not the schedule; update the schedule when it values the asset below market.
- Current $/token at $5.45 per GPU-hour: The Silicon Data B200 Rental Price Index, SDB200RT, shows a like-for-like B200 rental rate of $5.45 per GPU-hour. SemiAnalysis estimated that serving DeepSeek R1 on a B200 at 76 tokens per second per user costs $0.20 per 1 million tokens. Feed that rate into your cost-per-token model and replace any older rate.
- Buy-vs-rent break-even with 30-40% upfront: Compute providers are pushing AI startups toward three-year compute commitments with 30-40% down payments, replacing the earlier one-year reservation pattern. Test break-even against that term sheet before you sign; large upfront payments consume cash before tokens are served.
- 50-80% rental inflation stress test: B200 hourly rental prices moved from under $5 in January 2026 to $5.50-$5.80 by August 2026, a 50-80% increase over about seven months. Stress the next budget against that range; use the top of the range in the stress case.
- Renewal leverage from appreciation: Resale value above book value strengthens your exit option. Ask for a renewal quote against the current residual value and compare it with the new commitment. Quotes worse than the market call for a reset.
Break-even is a moving target
The old buy-versus-rent comparison ignored exit value. Now the break-even point moves because exit value is part of cost: buying lets the residual premium reduce effective cost, while renting lets the provider price appreciation into the rate. Keep both sides visible in the same model.
The stress test matters because rental rates can move quickly. Budgets built on a spot rate are fragile; budgets built on a range are defensible. Apply the observed range to the stress case, then check whether margin survives the top of the range. The margin check should cover power, network, and labor, not just the GPU line, because serving cost includes more than the GPU line.
Larger upfront payments are bets on the asset's exit value. Rising exit values make the bet easier to defend; falling exit values make renting more attractive.
Renewal leverage is the exit
Appreciation changes the negotiation. Providers that expect you to be locked in for a long term have less reason to offer a better renewal if you can show a stronger resale path. Keep the renewal quote next to the current market rate during review. Re-open the term sheet before signature when the quote is worse than the market.