Chip Economics
Earning power per spec-sheet unit.
Analytical decomposition of the published daily rates: derived analysis.
01 Four Lenses on One Rate
what the chip can hold — the size of model and context it can host
how fast it moves data — the dimension inference workloads buy
how fast it does arithmetic — training speed
what it costs to keep lit — rent per watt against facility cost per watt
Neocloud (T2) guaranteed on-demand, all regions, generations on age (x, years) against each lens (y, from zero). Spread = (max − min) ÷ mean across the five generations. Example prints · 2026-09-28 · rates from the 2026-09-28 snapshot. Headline stat per published rule (operator-equal median). The compute lens is denominated in petaFLOPs (PFLOP, 10^15 floating-point operations per second), dense, in the BF16 number format. Blackwell chips can also run FP4, a 4-bit format that Hopper and Ampere chips cannot use. For inference that runs in FP4, this lens understates what newer chips deliver per dollar.
One observed number underlies every panel: the Neocloud guaranteed on-demand rate, in dollars per GPU-hour. Each panel divides that rate by one spec-sheet denominator. One of the four compresses the cross-section far more than the others: memory bandwidth, at a 27% spread. Power spreads 32%, capacity 47% and compute 100%. Power is a looser second.
| Generation | Age | $/GPU-hr | n | $/GB-hr | $/TB/s-hr | $/PFLOP-hr | $/kW-hr |
|---|---|---|---|---|---|---|---|
| B300 | ~0.9y | $7.73 | 10 | 0.0286 | 1.004 | 3.43 | 7.03 |
| B200 | ~1.5y | $6.85 | 9 | 0.0381 | 0.890 | 3.04 | 6.85 |
| H200 | ~2.2y | $4.14 | 14 | 0.0294 | 0.862 | 4.18 | 5.91 |
| H100 SXM | ~3.8y | $3.56 | 21 | 0.0444 | 1.061 | 3.59 | 5.08 |
| A100 80GB | ~5.8y | $2.30 | 11 | 0.0287 | 1.128 | 7.37 | 5.75 |
CRI-T2-{chip}-SXM-GTD-OD-ALL-{GB|BW|PF|KW} · derived CRI series carry the identifier of the underlying cell in the first seven segments; strip the lens suffix to reach the parent. $/GB-hr computed from vendor high-bandwidth memory (HBM) capacity (80 / 80 / 141 / 180 / 270 GB).
02 The Bandwidth Band
Re-denominated per unit of memory bandwidth, datacenter silicon from 0.9 to 5.8 years of age rents inside $0.862–1.128 per TB/s-hour, a 27% spread around a $0.989 mean. Raw, the same five generations span 3.4×. The band is the test: a spread that widens and stays wide is the falsifiable marker firing. Members are disclosed on every exhibit.
The compression is a like-for-like property. It holds across cells that share an operator segment, interruptibility grade, term, and region: the series named under each exhibit. Pooled cuts, marketplace minimums, or mixed interruptibility grades will not reproduce it: grade and tenor are different markets per unit of bandwidth, as everywhere else on this site.
Power is the next tightest lens, at 32%. Rent per watt sorts by cooling class, which §03 shows.
CRB-T2-BW-ALL-OD-ALL · band row fields: band_min · band_max · band_mean · band_spread_pct · n_chips · member_chips. Emission gate n_chips ≥ 3, all members publication-qualified. Research-grade at launch (Shadow → Provisional).
The two parity rules. The sharpest way to see what the band rules out is to price the oldest member both ways. If compute were the priced unit, every new generation would force the prior one down to its share of the throughput: against the H100 print of $3.56, a chip with 32% of the BF16 throughput reprices to $1.12. If bandwidth is the priced unit, the same H100 print implies $2.16: the A100 carries 61% of the bandwidth. Two rules, two implied rates:
| Pricing rule | Assumes the market pays for | Implied A100 rate |
|---|---|---|
| Compute parity | BF16 throughput: 312 vs 989.5 TFLOPS | $1.12 |
| Bandwidth parity | HBM bandwidth: 2.039 vs 3.35 TB/s | $2.16 |
The observed A100 print is $2.30: 106% of bandwidth parity, 105% above compute parity, at 5.8 years of age. The market prices the A100 near its bandwidth share. That is the band's claim restated as a spread you can falsify: the day prior-generation prints migrate from the bandwidth-parity line toward the compute-parity line is the day the band breaks. The schedule to watch it on is the bandwidth frontier itself, which has moved more slowly than the compute frontier: the band's newest member entered at its predecessor's 7.7 TB/s.
03 Watt-Rent by Cooling Class
Per unit of nameplate power, the panel sorts by cooling class rather than by age. Air-cooled datacenter silicon rents at $4.49–5.75 per kW-hr (L40S, A100, H100). Liquid-class Blackwell rents at $6.85–7.03. H200, which runs in dense-air or liquid halls, sits between them at $5.91. Consumer cards on marketplaces, with no datacenter bundle, rent at $0.60–1.12. Within each class the prints sit close together, and each class sits on its own level. Watt-rent prices the hall a chip can occupy.
table view — watt-rent panel
| Generation | TDP | $/GPU-hr | n | $/kW-hr | Cooling class |
|---|---|---|---|---|---|
| B300 | 1.1 kW | $7.73 | 10 | 7.03 | liquid |
| B200 | 1.0 kW | $6.85 | 9 | 6.85 | liquid |
| H200 | 700 W | $4.14 | 14 | 5.91 | dense air / liquid |
| H100 | 700 W | $3.56 | 21 | 5.08 | air |
| A100 | 400 W | $2.30 | 11 | 5.75 | air |
| L40S | 350 W | $1.57 | 9 | 4.49 | air |
| 3090 (Mkt) | 350 W | $0.21 | 4 | 0.60 | consumer |
| 4090 (Mkt) | 450 W | $0.49 | 4 | 1.10 | consumer |
| 5090 (Mkt) | 575 W | $0.65 | 4 | 1.12 | consumer |
CRB-T2-KW-AIR-OD-ALL · CRB-T2-KW-LIQ-OD-ALL · cohort = cooling_class (air | liquid); consumer prints are Marketplace-segment context, not band members.
04 Economic Life: the Breakeven Altimeter
Watt-rent and cash operating cost are the same unit. A chip exits economic life when its watt-rent decays to the cash boundary beneath it. One log-scale strip therefore reads as an altimeter: today's rent bands above, the cash operating boundary below, and the gap between them is distance to shutdown. Air-cooled silicon today covers its cash operating cost ≈6–10×. The consumer band is the visible preview of late life, close above the boundary.
Worked illustration: hypothetical; every input elected by the reader
Elect a cash operating boundary c = $0.68/TDP-kW-hr (inside the cash band above) and a decay election g = 21%/yr (the cross-generation pace; see the caveat below). Air-band watt-rent today ≈ $5.12/TDP-kW-hr.
years to crossing = ln(watt-rent ÷ c) ÷ g = ln(5.12 ÷ 0.68) ÷ 0.21 ≈ 10
This is arithmetic on today's band against two elected inputs: a band-implied rent set against an elected cash cost. It is not a projection of any rate, and CCIR publishes neither election.
The honest gap. The −21% per year slope is a cross-generation reading: how fast rent falls with age across five generations on one date (/research/gpu-age-curve). It is not the band's own decay through time, which is the parameter the crossing needs. This page does not measure that decay. The band is also an observed equilibrium, not a law. In the 2023 shortage, frontier chips were posted far above any spec-sheet line. The flatness carries a date, which is the reason to keep watching it.
05 Method & Caveats
Normalized values equal the published parent series value divided by the disclosed constant. The parent cell's pooling, stat, and gate decisions are inherited unchanged. Denominators are vendor nameplate constants from NVIDIA datasheets, linked in the table below. A correction is a changelog entry, not a series break. The parent cell's published headline statistic (the operator-equal median) carries through unchanged.
| Lens | Denominator | A100 80GB | H100 SXM | H200 | B200 | B300 | Spec source |
|---|---|---|---|---|---|---|---|
| Capacity · $/GB-hrwhat the chip can hold: the size of model and context it can host | HBM capacity (GB) | 80 | 80 | 141 | 180 | 270 | NVIDIA datasheets |
| Bandwidth · $/TB/s-hrhow fast it moves data: the dimension inference workloads buy | memory bandwidth (TB/s) | 2.039 | 3.35 | 4.8 | 7.7 | 7.7 | NVIDIA datasheets |
| Compute · $/PFLOP-hrhow fast it does arithmetic: training speed | dense BF16 (PFLOPS) | 0.31 | 0.99 | 0.99 | 2.25 | 2.25 | NVIDIA datasheets · dense, not sparse |
| Power · $/kW-hrwhat it costs to keep lit: rent per watt against facility cost per watt | TDP (kW) | 0.4 | 0.7 | 0.7 | 1.0 | 1.1 | NVIDIA datasheets |
Denominators are variant-level: H100 SXM 3.35 TB/s and H100 PCIe 2.0 TB/s are different denominators. A chip without a disclosed denominator emits no row for that lens; nothing is imputed. Sources: A100 80GB · H100 SXM · H200 · B200 · B300: NVIDIA Blackwell Ultra datasheet, HGX B300 column (1,100 W TDP, 270 GB, 7.7 TB/s).
Lens admission rule
A lens is admitted when its denominator is a disclosed vendor nameplate constant, it answers a question a credit reader has, and it tells a distinct story. Considered and excluded on record: token throughput (measured, not nameplate), interconnect / fabric (a cluster property with no per-chip denominator; the tier axis prices the fabric), PUE-adjusted power (PUE is the reader's election), and FP8 / FP4 precision variants (a precision election; disclosure, not columns).
Blind spots.
- Denominators are nameplate constants. They capture none of a deployment's realized capacity, interconnect domain, or measured throughput.
- System draw exceeds chip TDP (1.82× for the DGX H100, per NVIDIA's rated 10.2 kW). TDP is the disclosed lens constant, and delivered-power cost comparisons apply the system factor.
- PUE, facility cost, and any decay election are the reader's. CCIR publishes prices, not the elections.
- The cross-section is one date across generations, not a cohort through time. §04 states the consequence.
- All numerators are posted list asks, not transactions. Thin cells (n < 3) are indicative.
Full construction: /documents/methodology · underlying cells: /explorer · the dated study these lenses generalize: /research/gpu-age-curve.