Every seller here is also a buyer
No spec sheet publishes what each of these companies buys. Runpod buys capacity from data-center partners and vetted hosts; it deployed three times its previous total fleet in Q1 2026 alone. DeepInfra buys data centers outright. AI Fabrik buys grid access before anything else, 150 MW under contract through GridCARE with connection in months rather than years, then fills modular edge sites with B300s. Lightning AI routes demand into seven or more third-party clouds while operating its own fleet. Firmus buys renewable power in hundred-megawatt blocks. Radiant goes one layer deeper. It buys permit-stabilized land banks, builds behind-the-meter generation, and racks NVIDIA systems on top, which makes it a power generator rather than a reseller. Vast.ai sits at the opposite limit: it owns no GPUs and buys no capacity at all. Over 1,400 independent providers list hardware they already paid for, and Vast now sells them the operating system: sourcing, financing against platform earnings, presold enterprise demand. Telnyx buys GPUs outright, racks them inside its own carrier PoPs, and buys no cloud capacity.
Knowing what a company buys tells you who its natural counterparty is, and deals happen there. The matrix below shows both sides.
Both sides of the sheet
| Company | Unit sold | Unit bought | Silicon | Scale signal | Center of gravity | Contract shape | Natural buyer | The bet |
|---|---|---|---|---|---|---|---|---|
| Telnyx | AI minutes + tokens | GPUs outright, carrier interconnects | Owned fleet (models undisclosed) | 4,000+ owned GPUs; ~$2.2M ever raised | 18 PoPs: US, EU, APAC | Per-minute ($0.08) · per-token, no minimums | Voice-AI & agent builders | The carrier edge beats the cloud edge |
| FriendliAI | Tokens, custom models | GPU capacity across clouds | NVIDIA via partners (B300) | 6–7× rev growth; 25–30 enterprise clients | US · Korea / APAC | Per-token · dedicated · BYO-cluster | Enterprises with custom models | Inference is a software problem |
| DeepInfra | Tokens | Data centers, GPUs | NVIDIA Blackwell, Rubin | ~5T tokens/week | 8 US data centers | Per-token · DeepCluster 3–5yr | App builders / agents | Agentic inference eats compute |
| AI Fabrik | Tokens, delivered at the edge | Grid access (150 MW via GridCARE), B300s, modular sites | NVIDIA B300, thousands deploying | First of 5 sites live Jul 2026; incubated in Gruve; Mayfield-backed | US tier-1/tier-2 metro edge | Undisclosed · enterprise + model cos | Latency-sensitive production inference | Tokens follow the CDN path |
| Runpod | GPU-seconds | Partner + host capacity | NVIDIA + consumer | ~$240M ARR; 500K devs | 31 global regions | Per-second, no commitment | Long-tail devs & startups | Long tail > whales |
| Lightning AI | Workflow + multi-cloud hours | Demand routed to 7+ clouds | Multi: fleet H100/B200/GB300 | 35K+ GPU fleet post-merger | US fleet, global marketplace | Seats + on-demand + reserved | Enterprise ML teams | Workflow > any single fleet |
| TensorWave | AMD GPU-hours | AMD silicon, 2GW+ capacity | AMD only, MI300X→MI355X | $493M raised; 8,192 MI325X live | North America | Reserved clusters | Memory-bound inference platforms | Memory/$ wins inference |
| Vast.ai | GPU-hours, marketplace | Nothing; supply is listed rather than bought | Host-owned, RTX 4090 → H200 | 17K+ GPUs · 1,400+ providers (Feb 2026); 98.6% B200 utilization (Jul 2026); ~$4M ever raised | 500+ locations, the global long tail | On-demand · interruptible auction · reserved, per-second billing | Cost-sensitive researchers & batch jobs | Liquidity beats ownership |
| GMI Cloud | Clusters + sovereignty | Power, DCs, hardware | NVIDIA Blackwell → Rubin | $12B Japan; ~7K GB300 Taiwan | Taiwan–Japan–US | Reserved + on-demand | APAC inference & enterprise | Sovereignty > spot market |
| Corvex | Clusters, confidential | NVIDIA allocation (NCP), DC capacity & power | NVIDIA H200/B200/GB200 → Rubin | Nasdaq: MOVE (Mar 2026); 1st verified CC on HGX B200 | US Tier III+ DCs · hybrid + on-prem | Multi-year reserved · hybrid EKS nodes | Model builders & regulated enterprises | Sovereignty by cryptography rather than geography |
| Firmus | MW / AI factories | Power, land, GB300s | NVIDIA GB300 → Rubin | $10B debt + $1.35B equity; 1.6GW target | Australia / APAC | Multi-year MW deals | Hyperscaler / sovereign | Green tokens earn a premium |
| Radiant | AI factories + GPU cloud | Powered land, on-site generation, NVIDIA systems | NVIDIA Blackwell → Rubin (DSX design) | $100B BAIIF pipeline; 5GW live / 45GW access claimed (Feb 2026) | Global land bank · sovereign-first | Long-term contracts + Ori cloud on-demand | Sovereigns, telcos, select enterprises | Compute is a utility; capital cost wins |
The public rows earn the map. The private rows form the ledger, shared case-by-case with each provider's consent.
Ask about a specific provider →The deep cards
$10B Blackstone-led debt, 1.6GW Project Southgate, and a bet that green tokens earn a pricing premium.
Read the card → Rung 1 · MW, the utility modelRadiantBrookfield's compute vehicle: the Ori merger (Feb 2026), a $100B BAIIF pipeline, and claimed access to 5GW live power: the AI-factory-as-utility bet. Draft, pending verification.
Read the card → Rung 2 · Clusters + sovereigntyGMI Cloud~7,000 GB300s in Taiwan at near-full utilization and a $12B sovereign initiative in Japan.
Read the card → Rung 2 · Clusters, confidentialCorvexThe security-first GPU cloud, now public (Nasdaq: MOVE); first verified confidential computing on HGX B200, model weights invisible even to the host. Draft, pending verification.
Read the card → Rung 3 · GPU-hours, AMDTensorWaveThe all-AMD cloud with Fireworks AI and Luma AI in production: the ROCm objection, answered with evidence.
Read the card → Rung 3 · GPU-hours, marketplaceVast.aiThe pure marketplace: 17K+ GPUs from 1,400+ providers in 500+ locations, on ~$4M ever raised; 98.6% B200 fleet utilization, and now selling hosts the OS too: sourcing, financing, presold demand. Provider-reviewed.
Read the card → Rung 4 · Workflow + hoursLightning AIThe PyTorch-native platform that just became supply: 35,000+ GPUs after the Voltage Park merger.
Read the card → Rung 5 · GPU-secondsRunpod~$240M ARR on ~$22M raised before its June round, and the biggest capacity buyer of the six.
Read the card → Rung 6 · TokensDeepInfra~5T tokens a week, ~30% from agents: the clearest public signal that agentic workloads are infrastructure-scale.
Read the card → Rung 6 · Tokens, custom modelsFriendliAIThe continuous-batching inventor's asset-light answer to DeepInfra: sells tokens from your model, buys GPUs across clouds. Draft, pending verification.
Read the card → Rung 6 · Tokens, delivered at the edgeAI FabrikThe Gruve spinout building a CDN for tokens: modular sites near US metros, 150 MW under contract via GridCARE, first of five live July 2026. Draft, pending verification.
Read the card → Rung 7 · AI minutesTelnyxThe carrier that became an inference provider: 4,000+ owned GPUs inside 18 telephony PoPs, selling conversation minutes at $0.08 all-in, on ~$2.2M ever raised. Draft, pending verification.
Read the card →Frequently asked
What is the compute ladder?
A classification of AI infrastructure providers by the unit of consumption they sell, not the hardware they run: megawatts → AI factories → GPU clusters → GPU-hours → GPU-seconds → tokens → AI minutes. Buyers should shop the rung that matches their team's abstraction level.
What's the difference between a GPU cloud and an inference provider?
A GPU cloud sells time on hardware the buyer must operate; an inference provider sells the output, tokens through an API, and hides the hardware. Companies now span rungs: DeepInfra sells tokens and dedicated clusters; Lightning sells a platform, its own fleet, and third-party capacity.
Which of these companies also buy capacity?
Runpod (from DC partners and vetted hosts), DeepInfra (data centers outright), and Lightning AI (demand routed to 7+ clouds). The "both-sided" companies are the most active counterparties in the market; they transact in two directions.
Why these companies, and why is the top of the ladder crowded?
We add companies rung by rung so the ladder spans every unit, and contested rungs show every posture: around tokens, DeepInfra owns the metal, FriendliAI buys capacity, AI Fabrik sells proximity from edge sites near US metros, and Telnyx owns the metal and the network, selling minutes above it; around clusters, GMI Cloud sells sovereignty by geography while Corvex sells it by cryptography; on GPU-hours, TensorWave owns an all-AMD fleet while Vast.ai owns no GPUs at all, a 1,400-provider auction; on megawatts, Firmus sells a renewable premium while Radiant sells Brookfield-scale capital and powered land. The full market lives on the ecosystem map; cards ship as we verify them.
How these cards are built
Every public figure carries a date and traces to a primary source: company newsrooms, funding press releases, filings, and named-customer announcements. We re-verify cards on update and stamp them. The five private-ledger fields come directly from providers in conversation and stay unpublished; we share them case-by-case with the provider's consent. Spot an error or a stale number? Tell me; corrections ship within a week.