CRI-ONE Essay · 2026-09-10
Photons per Dollar
Where the market prices compute in 2026 — from the Landauer thermodynamic floor to the top of the published per-token ceiling — and why every number on cri-one.com sits exactly where it sits.
I. The question
Somebody clicks BUY. A number lands on their card. What did they just buy? For a physical good, the answer is easy: some atoms in a specific arrangement, weighed and shipped. For compute — a FLOP, a token, a photon-across-a-fiber — the answer is stranger. You bought a small, precise release of energy from a chip and its power supply, arranged so that the outputs mean something. The dollar you spent is a proxy for that energy, that meaning, and the cost of everything the vendor put around it: silicon, cooling, real estate, engineering, insurance, profit.
The question this essay tries to answer is narrow: for that one FLOP, what does the market currently ask? Where is the floor set by physics? Where is the ceiling set by scarcity? And where does an enterprise catalog like CRI-ONE sit inside that bracket — not by feel, but by number?
II. Three units, one commodity
The industry sells the same commodity three different ways, and the confusion of units is the first thing to clear.
$/hr for a GPU or VM. This is what AWS, Google, and Azure quote for renting silicon. A NVIDIA H100 on AWS is $6.88 per GPU-hour list, $12.29 on Azure, $4.29 on Lambda Labs, and as low as $0.35 on the Vast.ai spot marketplace. The forty-fold spread inside the same chip is entirely about who buys the reliability guarantee.
$/token for an LLM API. This is what OpenAI, Anthropic, Google, xAI, Together, and Fireworks quote for hosted inference. In 2026 the frontier band runs $5–$50 per million output tokens (Claude Opus 5 at $25, GPT-5.6 Sol at $30, Fable 5.1 at $50, Gemini 2.5 Pro at $10). The reasoning ceiling for anything published per-token is OpenAI o3-pro at $80/1M. Above that, the industry stops selling by the token and starts selling by the month.
$/FLOP or W/FLOP. This is what chip architects actually think about. Cost per floating-point operation, and energy per floating-point operation. Both are strictly bounded from below by physics. Both are the only unit that lets you compare a hyperscale GPU to a spot-marketplace GPU to an LLM API to a photonic tape-out program — because in the end all four are turning joules into meaning.
Compute is priced three ways because three different buyers ask three different questions. But underneath, it is one commodity: energy, arranged as answer.
III. The floor
Physics sets a hard, unarguable lower bound on the cost of a FLOP. In 1961, Rolf Landauer proved that erasing one bit of information in an irreversible logic gate at temperature T must dissipate at least kT · ln 2 joules of heat. At room temperature (300 K), that works out to 2.87 × 10−21 J per bit erased. Convert to dollars at U.S. retail electricity ($0.10/kWh = $2.78 × 10−8/J) and the thermodynamic floor for a bit erasure is roughly 8 × 10−29 dollars.
A modern floating-point operation touches on the order of a few hundred bits internally — multiply, add, normalize, register write. Take a mid-range estimate of 64 bit-erasures per FLOP and the Landauer floor becomes ~5 × 10−27 dollars per FLOP. That is the number below which no irreversible processor at 300 K can go, ever, at any scale, on any budget. Reversible-computing designs (Bennett 1973) could in principle undercut it, but no commercial reversible processor exists in 2026.
Photonics has its own floor, from the quantization of light itself. A photon at the 1550 nm telecom wavelength carries exactly hc/λ = 1.28 × 10−19 J of energy. To transmit one bit optically you must deliver at least one photon, so every optical compute path must burn at least that much energy per bit. Photonics wins on data movement, not on approaching Landauer.
IV. The frontier
Modern silicon lives about six to eight orders of magnitude above the Landauer floor. A NVIDIA H100 SXM at BF16 dense throughput runs at 0.35 pJ per FLOP — that is, 3.5 × 10−13 joules. Multiply by retail electricity and one H100 FLOP costs about 1 × 10−20 dollars of electricity. Compare against the Landauer 64-bit floor (~5 × 10−27 $/FLOP) and modern silicon sits about a million times above the physics limit — almost all of the energy goes into wire and gate charging, not the logical erasure itself.
The 2026 frontier has moved. Blackwell B200 in FP4 dense hits 0.111 pJ/FLOP. Microsoft Maia 200 gets to 0.075 pJ/FLOP at FP4. AMD MI355X approaches 0.070 pJ/FLOP. Google's Trillium (TPU v6e) at $2.70/chip-hour gives the best hyperscaler-native $/FLOP of the year: about 8.2 × 10−19 $/FLOP-hour. Every current tensor engine is already sub-1 pJ/FLOP at native precision; the 2026 race is sub-100 fJ/FLOP at FP4.
V. The market
Convert per-token API rates into per-FLOP terms (assuming a 70B-class model burns roughly 3.5 × 1012 FLOPs per output token) and the LLM market sits in a tight band on the log scale. DeepInfra Llama 3.3 70B at $0.12/1M output is about 3.4 × 10−20 $/FLOP. Claude Opus 5 output at $25/1M is 7.1 × 10−18 $/FLOP. OpenAI o3-pro at $80/1M output — the highest per-token API rate published anywhere — is 2.3 × 10−17 $/FLOP. Above that, capacity is only sold reserved: Fireworks dedicated H100 at ~$7/hr, Anthropic enterprise at ~$441k/yr average, OpenAI enterprise at ~$318k/yr average, Quantinuum quantum subscription at $125k–$175k/month.
Everything published for public purchase in 2026 fits in the same three-decade band: from about 10−20 $/FLOP at the cheap open-weight floor to about 10−17 $/FLOP at the o3-pro ceiling. Nothing sold by-the-token exceeds that.
VI. Where CRI-ONE sits
The RIDDLE ladder is calibrated against these market points, deliberately. Every tier's copy names the market comparable so a technical buyer can see the reference. The prices are NDA-gated in public display, but the anchor points are the ones below.
| Tier | Anchor rate | Nearest market comparable | $/FLOP-equiv |
|---|---|---|---|
| P1 Baseline | $8/1M | Claude Opus 5 input $5, GPT-5.6 Sol input $5, Sonnet 5 output $10 | 2.3×10⁻¹⁸ |
| P2 Reserved Premium | $32/1M | Claude Opus 5 output $25, Fable 5.1 output $50, Grok 4.6 output $30 | 9.1×10⁻¹⁸ |
| P3 Top per-token band | $128/1M | o3-pro $80, GPT-5 Pro $120, GPT-5.2 Pro $168 (public ceiling) | 3.7×10⁻¹⁷ |
| P4 Reserved capacity / SLA | $512/1M | Azure PTU at ~30% utilization ($500–$1100/1M output) | 1.5×10⁻¹⁶ |
| P5 Contact-for-quote | $1,995/1M | Quantinuum H2 subscription $125k–$175k/mo; OpenAI enterprise quote-only | 5.7×10⁻¹⁶ |
P1–P2 sit inside the frontier band. P3 is the top of the published per-token map. P4 crosses out of pay-per-token and into reserved-capacity territory. P5 leaves the token map entirely and sits with the enterprise-subscription tier where things stop being commodities and start being contracts.
The QBeam die catalog is a different animal. Twelve die generations from APM01 (7 nm, 50–125 Gbps) to APM12 (2 nm, 2.4–3.4 Tbps aggregate). List prices from $10M for APM01 to $250M for APM12, structured as IP-license + tape-out packages rather than per-unit silicon. The frame of reference is not merchant chips (Cerebras CS-3 wafer-scale at ~$2–3M, GB200 NVL72 rack at ~$3M) but bespoke photonic-ASIC dev programs: Lightmatter Envise/Passage enterprise engagements, PsiQuantum's $1B utility-scale contracts, Amazon Trainium generational programs. In that peer set, an IP license + first tape-out + N wafers at $25M–$250M is inside the credible range for a signed engagement.
Every price on cri-one.com is a specific point in a six-decade physics-to-market band, with a specific market comparable named next to it. The number is not chosen; it is placed.
VII. How you buy compute in 2026
The last change is procedural, not substantive. In 2026 you do not wait for a countersigned PDF. Every enterprise button on cri-one.com is a click-through electronic signature under the U.S. E-SIGN Act (15 U.S.C. § 7001) and UETA: check the acceptance box, click Start, and the mutual NDA + wire instructions + owner alert fire in the same second. The buyer's timestamp, IP, and user-agent hash are recorded as the signature witness. For purchases under $999,999.99, Stripe Checkout carries the charge directly; over that cap, wire is mandatory and the wire memo line arrives in the same click. No paper, no round-trip.
That is what the essay is about, in the end. Not the number on the card, but the alignment: physics floor at 10−27 $/FLOP, market ceiling at 10−17 $/FLOP for anything sold by the token, six decades of defensible territory in between, and a click-through agreement so the moment the buyer commits, everything fires at once. When someone hands over money in that arrangement, the number they see equals the number physics and the market both agree they should see. That is what "earned" means, priced in FLOPs.
w7n17s2sp and w5j3cs2jn, 2026-09-10.
— Christopher Gabriel Brown / CRI-ONE