A spot market for LLM inference · product of Spectral Labs (Boltzmann Ltd) · thegrid.ai
Real engineering, genuine invention, zero external validation, and a raise that disqualifies it before any of that matters. The product exists and runs. The company behind it has raised $29.75M, which trips our $15M hard filter. And every claim of market adoption traces back to catalog entries the company filed itself.
thegrid.ai is not a separate company. Spectral Labs’ own site
features The Grid as its active platform and offers SPEC token holders “$50 in credits on The Grid.” The legal entity in
Spectral’s footer is Boltzmann Ltd. The CEO’s only current role in Harmonic is Spectral Labs. Five of nine Grid
staff hold current Spectral roles. Spectral’s headcount fell 27 → 17 while Grid’s rose
7 → 13. This is one team moving products, not a new cap table.
Total raise $29.75M > $15M → excluded before scoring.A continuous limit order book for LLM output. You do not buy a model, you buy an instrument: task type (Text / Code / Agent) × tier (Standard / Prime / Max). Each tier is a contractual quality floor — thresholds on Artificial Analysis benchmark indices plus p90 latency and throughput ceilings. Suppliers post offers on the same instrument and the cheapest qualifying offer wins each request. Advanced mode adds limit orders, positions, and resale of unused tokens back onto the book. Published prices: $0.035 / $0.120 / $1.709 per 1M tokens.
The trade you are asked to accept: “the specific model behind any given call can change between calls.”
api.thegrid.ai/v1/models returns 200 with 17 real instruments.trading.api.thegrid.ai/v1/health returns ok; /v1/instruments returns a genuine 401 auth wall.pricing is null for all 17 models on the public API, so it cannot be audited from outside./terms, /privacy, /legal all 404. No entity in the footer. Their own published LinkedIn URL 404s.Their docs present fifteen tools as integrations. None is confirmed by the counterparty. All of it descends from a single community catalog entry the company filed itself.
| Claimed | Counterparty confirms? | What is actually there |
|---|---|---|
| LiteLLM | No | PR by Grid staff closed, not merged; a second PR still open. Zero mentions in their docs. |
| Cline | No | One auto-generated line whose own field reads “from models.dev”. Their doc page 404s. |
| Kilo Code | No | A single string in an alphabetical icon-name list. Zero mentions in 68KB of docs. |
| Cursor, Continue, OpenCode, CrewAI, DeerFlow, Manifest | No | Zero mentions across all of their documentation. Pull-through from models.dev only. |
| “Hermes Agent” | No | A Grid staffer forked it and wrote the integration doc himself. Upstream never mentions The Grid. |
| AgentSea | Their own product | Lives inside The Grid’s own GitHub org. Listing it beside real third parties is misleading. |
| OpenRouter | No | Not listed on OpenRouter at all. |
models.dev, litellm and catwalk on one day and filed provider PRs into each. Catalog
seeding is a legitimate go-to-market tactic. Presenting it as adoption is not.| Who | Raised | Why it hurts |
|---|---|---|
| OpenRouter | $164.2M | :floor already routes to the cheapest qualifying provider. 400+ models, 94 staff, a16z / Sequoia / CapitalG / Alphabet / NVIDIA. |
| Azure AI Foundry model router | Microsoft | “Cost mode” sells a quality band and picks the cheapest model clearing it. That is The Grid’s pitch, shipped by Microsoft, free, enterprise-governed. |
| Vercel AI Gateway | Vercel | Hundreds of models at explicitly zero markup. Hard to charge a spread against zero. |
| LiteLLM | free / OSS | routing_strategy: cost-based-routing. About 20 lines of YAML, and you keep knowing which model answered. |
| Together / Fireworks / Baseten / CoreWeave / DeepInfra | $153.6M – $28.8B | They are the supply and the competition, and they reach the same buyers directly. |
| Silicon Data | $35.2M | Already builds the token price index any commodity market settles against — backed by Jump, DRW, Wintermute and distributed on Bloomberg and Refinitiv. |
The exchange layer is genuinely empty and would be winner-take-most if it ever worked, because liquidity is the only real network effect here. But it is empty because demand has not appeared, not because it is hard — and to reach it The Grid must first win a price-only fight in the most competitive, lowest-margin, best-capitalised segment of AI infrastructure, against free.
Benchmarks used: Auron Solutions 4.0 IS / 6.25 CAN-BE (had signed GDMS and Chevron agreements — The Grid has no counterparty confirmation of anything). Oligo Space 3.5–4.5 IS (token/SPV wrapper, same structural problem). Lamb Labs 3.0 (no product at all — The Grid is clearly above it). Delivr 3.5 (numbers that did not reconcile). Cosmic Shielding 6.71 and Satlyt 6.75 sit above on real revenue and real customers.
Recommendation: pass. The filter decides it. $29.75M behind the team, a token rather than equity, and a Daxos check that would be a rounding error on a cap table we cannot see.
Reopen only on a clean answer to one question: is The Grid a separate company with its own equity cap table, or is it Spectral Labs wearing a new name? If genuinely separate, sub-$3M, in equity, and the price arithmetic reconciles — this becomes a real conversation at 6.0. Ask for the incorporation certificate, the cap table, and the take rate. Everything else follows from those three documents.