Two providers rent you the same B200. One charges 1.38x its own H100 price for it. The other charges 2.38x. Same silicon, same quarter, same country — and the gap between those two multipliers is the cleanest window we have into what a GPU-hour actually costs versus what scarcity lets providers charge.
On Lambda's on-demand tier, a B200 runs $5.50/GPU-hour against $3.99 for an H100 — 1.38x the price for 2.27x the FP8 (sparse) throughput. Per unit of effective compute, the newest part is about 39% cheaper than the old one. On AWS capacity blocks — the same-tier comparison on the hyperscaler side — a B200 runs $9.36 against $3.93 for an H100: 2.38x the price for the same 2.27x the compute. Per-PFLOP parity. One market prices performance; the other prices a queue.
This note decomposes the rented GPU-hour into its three constituent parts — hardware amortization, energy plus facility, and the residual demand premium — for the H100 SXM5, H200, and B200, using our own cost constants and the April 2026 US-region price snapshot. All derived figures are Silicon Analysts estimates; the math is shown so you can disagree with any input and recompute.
All rental prices in this note are US-region list prices as of 2026-04-07, taken from provider public price pages. Comparisons are always within a single pricing tier (on-demand vs. on-demand, capacity block vs. capacity block) — blended or cross-tier comparisons are deliberately excluded.
1. Methodology: splitting a GPU-hour three ways
Every rented GPU-hour decomposes into exactly three buckets:
Bucket 1 — Hardware amortization. We take the GPU street price (SA estimates: H100 SXM5 $28,000, H200 $38,000, B200 $40,000), apply a ~1.4x server-overhead multiplier to capture the host system the GPU cannot run without — chassis, CPUs, system memory, storage, NICs, in-node switching (SA estimate) — then amortize straight-line over a 4-year refresh cycle with a 30% end-of-life residual value, spread across utilized hours only:
hardware $/hr = (street price x 1.4) x (1 - 0.30) / (4 x 8,760 x utilization)
We use a 65-75% utilization band — the range spanning continuous training fleets (~65%) through well-orchestrated, containerized fleets (~75%) per industry utilization benchmarks — with 70% as the central case. At 70%, the denominator is 24,528 billable hours.
Bucket 2 — Energy plus facility. Electricity at the GPU's rated TDP, uplifted by a PUE of 1.12 (liquid-cooled AI datacenter), at Northern Virginia rates ($0.058/kWh — the largest US datacenter market): TDP x 1.12 x 0.058 / 1000. On top of that, a facility proxy: NoVA colocation at $207/kW-month applied to the GPU's TDP, converted at 730 hours/month, and divided by utilization since the space is paid for whether the GPU is rented or not. The blended energy-plus-facility figure is an SA estimate.
Bucket 3 — Demand premium (the residual). Observed same-tier rental price minus buckets 1 and 2. This is everything that is not hardware or facility cost recovery: provider gross margin, sales and support overhead, and — dominating both in this market — scarcity rent.
Cross-check on the 1.4x multiplier. Our AI server bill-of-materials dataset puts an 8-GPU H100 system at roughly $45,000 in total component cost against $26,600 of GPU content — a 1.69x system-to-GPU ratio at the cost level, or about $5,600 per GPU slot. At street prices the multiplier compresses, because the GPU carries a far higher margin than the balance of the system; 1.4x on a $28,000 H100 implies ~$313,600 for an 8-GPU server, consistent with reported HGX-class system pricing (SA estimate).
What's excluded. The B300 appears in AWS's capacity-block price ladder ($11.70/GPU-hour, waitlist) and we chart it below — but it has no entry in our chip specification database, so there is no street-price or TDP basis to decompose it. It is excluded from the decomposition rather than estimated. GCP's B200 is listed as spot-only in our snapshot ($6.69/GPU-hour derived from the 8-GPU instance); spot is a different tier, so it is excluded from same-tier comparisons. Azure's public price sheet in our April snapshot carries only two SKUs — H100 ($12.29/GPU-hour on-demand) and H200 ($13.78) — and no Blackwell listing at all.
2. The decomposition: most of the GPU-hour is premium
Running the three buckets at 70% utilization against representative neocloud on-demand prices:
At 70% utilization, the demand premium is the largest slice of every neocloud GPU-hour — 52-64% of the price.
Straight-line 4-yr amortization, 30% residual, 1.4x server overhead, NoVA power + colo. All components are SA estimates.
Source: Silicon Analysts estimate; derived from provider public price pages (Apr 2026), NVIDIA spec sheets, Uptime Institute / EIA benchmarks.
The full numbers, including where the reference price comes from:
| SKU | Reference price (tier) | Hardware $/hr | Energy + facility $/hr | Demand premium $/hr | Premium share |
|---|---|---|---|---|---|
| H100 SXM5 | $3.99 (Lambda, on-demand) | $1.12 | $0.33 | $2.54 | 63.7% |
| H200 | $3.89 (CoreWeave, on-demand) | $1.52 | $0.33 | $2.04 | 52.5% |
| B200 | $5.50 (Lambda, on-demand) | $1.60 | $0.47 | $3.43 | 62.4% |
All derived columns are SA estimates at 70% utilization. Components sum to the observed price by construction — the premium is the residual.
Three things stand out.
Electricity is almost irrelevant. An H100 at 700W TDP, PUE 1.12, NoVA rates burns about $0.045/hr of electricity — barely 1% of the rental price. A B200 at 1,000W burns $0.065/hr. The "GPUs are power hogs" narrative is true for grid planning and false for rental economics: even doubling the power price moves the cost stack by a nickel. The facility slice (colo space at $207/kW-month) is 4-6x larger than the electrons themselves. For the manufacturing-side view of energy costs, see our chip manufacturing energy cost analysis.
The H200's premium is the thinnest. At $3.89 on CoreWeave, the H200 carries only a 52.5% premium — because its hardware bucket is heavier ($38,000 street price vs. $28,000) while renting below the H100's neocloud price. CoreWeave lists its own H100 at $4.76 — an inversion where the strictly better part rents for 18% less on the same provider, same tier. That is what deliverability does to price: H200 supply caught up; on-demand H100 capacity at premium neoclouds still clears at scarcity prices. RunPod corroborates the level, listing H200 at $3.99 on-demand.
Hardware amortization sensitivity is modest within the realistic band. Across the 65-75% utilization range (SA estimates):
| SKU | Hardware $/hr @ 65% | @ 70% | @ 75% |
|---|---|---|---|
| H100 SXM5 | $1.20 | $1.12 | $1.04 |
| H200 | $1.64 | $1.52 | $1.42 |
| B200 | $1.72 | $1.60 | $1.49 |
Even at the pessimistic end of the band, the H100 cost stack tops out around $1.53/hr all-in — still under 40% of the Lambda price, and 12% of the Azure price. The conclusion — most of the GPU-hour is premium — survives any utilization assumption a competent operator would admit to. What the silicon itself costs to make is a separate question — our B200 Blackwell cost breakdown walks the manufacturing stack, and you can rebuild it interactively in the Chip Price Calculator or compare across accelerators in the Cost Bridge.
3. Per-unit-of-compute pricing: two regimes, one market
Divide the same-tier price by FP8 (sparse) throughput — 3,958 TFLOPS for both H100 SXM5 and H200, 9,000 TFLOPS for B200, per NVIDIA spec sheets (sparse figures; dense is roughly half) — and the market's two pricing regimes separate cleanly:
Lambda sells B200 compute at $0.61 per FP8 (sparse) PFLOP-hour — 39% below its own H100. AWS prices Blackwell at per-PFLOP parity.
Sparse throughput per NVIDIA spec sheets; dense is ~half. Derived $/PFLOP-hr are SA estimates.
Source: Silicon Analysts estimate; derived from provider public price pages (Apr 2026), NVIDIA spec sheets, Uptime Institute / EIA benchmarks.
On the neocloud side, Blackwell is a per-performance discount: Lambda's $5.50 B200 works out to $0.61 per effective FP8 (sparse) PFLOP-hour against $1.01 for its H100 — the 1.38x price buys 2.27x the compute. Hopper-class compute clusters around $1.00/PFLOP-hr regardless of provider tier: Lambda H100 on-demand $1.01, AWS H100 capacity block $0.99, CoreWeave H200 on-demand $0.98. That convergence is what a competitive market looks like.
On the hyperscaler side, AWS prices its B200 capacity blocks at $1.04 per FP8 (sparse) PFLOP-hour — 4.7% above its own H100 capacity block. Two years of silicon progress, zero per-performance discount. The H200 capacity block is worse still at $1.26 — a 27% per-PFLOP surcharge over H100 for the same compute plus more memory (141 GB vs. 80 GB), a premium AWS raised ~15% in January 2026 according to its published capacity-block prices. And hyperscaler on-demand tiers sit in a different universe entirely: Azure's on-demand H100 works out to $3.11 per FP8 (sparse) PFLOP-hour and GCP's to $2.77 — roughly 3x the neocloud level for identical silicon (all per-PFLOP figures SA estimates).
The interpretation is not that hyperscalers are inefficient. It is that they are not selling GPU-hours — they are selling guaranteed, scheduled access inside an ecosystem, and pricing the guarantee at what the queue will bear. Track how these curves move in our Cloud GPU Pricing tracker.
4. What the premium implies
The capacity-block ladder is a queue price
AWS's capacity-block ladder — the only place all four current NVIDIA datacenter SKUs appear at the same tier — is the purest scarcity signal in the dataset:
AWS capacity-block pricing steps 2.4x from H100 to B200 — and the waitlisted B300 tops the ladder at $11.70 before general availability.
Same provider, same tier. B300 is excluded from our cost decomposition — no public spec basis yet.
Source: Silicon Analysts estimate; derived from provider public price pages (Apr 2026), NVIDIA spec sheets, Uptime Institute / EIA benchmarks.
Note the shape: the ladder is not a cost curve. The B200's manufacturing cost is roughly double the H100's, its street price about 1.4x — but its capacity-block price is 2.38x. And the B300 enters the ladder at $11.70 while still waitlisted, 25% above the B200, before anyone can characterize its real-world throughput. Prices like that are set by queue position, not by bill of materials.
Premium share is a provider fingerprint
Expressing the demand premium as a share of the observed price, per provider and tier:
Same silicon, same math: the demand premium runs 52-64% of the price at neoclouds and 78-88% at hyperscalers.
Premium = observed price minus SA-estimated hardware + energy/facility stack at 70% utilization.
Source: Silicon Analysts estimate; derived from provider public price pages (Apr 2026), NVIDIA spec sheets, Uptime Institute / EIA benchmarks.
The neocloud cluster (52-64%) and the hyperscaler cluster (78-88%) do not overlap. On Azure's on-demand H100, $10.84 of the $12.29 hourly price is premium over the cost stack (SA estimate) — same silicon whose all-in cost stack we put at $1.45/hr. Even the cheapest US on-demand H100 in our snapshot — RunPod at $2.69 — still carries a ~46% premium (SA estimate). There is no tier in this market, at any provider, where GPU-hours rent near cost.
Reserved tiers are the forward curve
If the premium is scarcity rent, then the discounts providers offer for term commitments are their own forecast of how fast that scarcity decays:
| Provider | Spot discount | 1-yr reserved | 3-yr reserved |
|---|---|---|---|
| AWS | 50-90% off | 25-31% off | Up to 45% off |
| Azure | ~20-30% off | ~30% off | ~40-50% off |
| GCP | 60-91% off | ~30% off | ~55% off |
| CoreWeave | N/A | Up to 60% off | Up to 60% off |
Source: provider public pricing documentation, Apr 2026 snapshot.
Read this table against the decomposition: a 45-55% discount for a 3-year commitment prices the committed GPU-hour at roughly where the premium share compresses from ~85% toward ~65% — hyperscalers converging on today's neocloud economics, three years out. CoreWeave offering up to 60% off for terms tells you a neocloud with visibility into its own pipeline expects Hopper-class scarcity rent to largely evaporate inside the commitment window. The on-demand price is a queue price; the reserved price is the provider's actual supply forecast, with money behind it.
There is also direct evidence that premiums reprice administratively, not with costs: AWS cut on-demand H100 prices 44% in June 2025, then raised H200 capacity-block prices ~15% in January 2026. Nothing in the hardware or energy stack moved even a cent in either event — the premium did all the work. Weigh renting against owning with these same constants in our Build vs. Rent and Cluster TCO calculators.
5. Sensitivity: utilization is the biggest lever
Every input above is contestable. Only one of them materially moves the answer. Re-running the decomposition across a 50-90% utilization range (both the hardware and facility buckets scale with utilization; electricity does not):
Even at a punishing 50% utilization assumption, every SKU's implied premium stays positive — the scarcity-rent conclusion is not an artifact of the utilization input.
Premium = on-demand price minus SA-estimated cost stack, recomputed at each utilization level.
Source: Silicon Analysts estimate; derived from provider public price pages (Apr 2026), NVIDIA spec sheets, Uptime Institute / EIA benchmarks.
At 50% utilization — a fleet half-idle over its entire four-year life — the implied H100 premium is still $1.98/hr, half the Lambda price. At 90%, it is $2.85/hr. The premium's existence is robust; its size swings by roughly 45% across the range, which is more than any plausible disagreement about street prices, electricity rates, or colo rates can produce. If you take one number into your own model, make it your utilization assumption — and pressure-test it in the Cluster TCO calculator.
6. Methodology notes and limitations
The full assumption set, and where it bites (all items SA estimates unless sourced):
- Depreciation life is the second-biggest lever. We use a 4-year refresh cycle. On a 3-year life, the H100 hardware bucket rises from $1.12 to $1.49/hr; on 5 years it falls to $0.90/hr. Hyperscalers have publicly stretched accounting lives to 5-6 years; AI-native operators often model 3.
- Cost of capital is excluded. Straight-line amortization ignores the time value of money. Applying our 8% WACC default as a capital-recovery annuity raises the H100 hardware bucket ~34% (to about $1.50/hr) and shaves the premium accordingly — it does not change any qualitative conclusion.
- The 30% residual is generic, not SKU-specific. Our H100-specific residual curve puts a 4-year-old unit at 15% of original value, which would raise the H100 hardware bucket to about $1.36/hr. We use the fleet-level 30% default for cross-SKU consistency.
- The 1.4x server overhead is a street-price estimate. The component-cost ratio in our server BOM dataset is 1.69x; margin structure compresses it at acquisition prices. Reasonable alternatives (1.3-1.5x) move the hardware bucket by less than $0.10/hr.
- Energy is charged on GPU TDP only. Host CPUs, fans, and NICs add roughly 10-20% to real draw; conversely, our colo proxy (space priced on GPU TDP, plus separately metered electricity) likely overstates facility cost for owner-operated datacenters. The two biases partially offset, and both are small against the premium.
- Prices are US-region list, single snapshot (2026-04-07). Negotiated enterprise pricing sits below list; regional and availability-zone spreads exist. Directionally, negotiated discounts make the list-price premium an upper bound on what sophisticated buyers pay — not a lower one.
- FP8 throughput figures are sparse. All per-PFLOP figures use NVIDIA's sparse spec-sheet numbers (H100/H200: 3,958 TFLOPS; B200: 9,000 TFLOPS); dense throughput is roughly half, and real-world utilization of peak FLOPS varies by workload. Ratios between SKUs are less sensitive to this than absolute $/PFLOP levels.
- B300 is excluded from the decomposition. It appears in the AWS price ladder but has no entry in our chip specification database — no street price, no TDP — so any decomposition would be invention.
The one-paragraph takeaway
At 70% utilization, a fully burdened H100 costs about $1.45/GPU-hour to own and operate — $1.12 of straight-line hardware amortization and $0.33 of energy plus facility (SA estimate) — yet rents for $3.99 at Lambda, $4.76 at CoreWeave, and $12.29 on-demand at Azure. The residual demand premium runs 52-64% of the price at neoclouds and 77-88% at hyperscalers, and no tier at any provider in our April 2026 snapshot rents GPU-hours near cost. The market has split into two regimes: neoclouds price per unit of compute — Lambda's B200 at 1.38x its H100 price for 2.27x the FP8 (sparse) throughput is a 39% per-PFLOP discount — while hyperscalers price the queue, with AWS capacity blocks holding Blackwell at per-PFLOP parity and the waitlisted B300 entering at $11.70. Electricity, at a nickel an hour, is noise. The numbers to watch from here are not hardware costs — they barely move the stack — but the reserved-tier discounts (45-60% off for term commitments), because they are the providers' own money-backed forecast of how fast the scarcity rent decays.