{"recordType":"header","exportVersion":1,"exportedAt":"1790978632660","producer":{"desktopVersion":"0.0.0","buildCommit":"development","platform":"darwin","architecture":"arm64"},"conversation":{"id":"conversation:1","title":"Compute Markets","createdAt":"1790780269985","harnessConfigurationName":"codex-basic","permissionProfileId":"full"},"turns":[{"id":"turn:1","sequence":1}]} {"recordType":"turn","id":"turn:1","sequence":1,"createdAt":"1790780269985","text":"What is the current state of the art for compute markets, who are the players and what do experts in the field predict it could look like","interactionNodeId":"node:1","origin":{"kind":"user"},"completion":{"status":"accepted","attemptOutcome":"accepted","harnessConfigurationName":"codex-basic","modelSelection":{"providerId":"codex","modelId":"gpt-5.6-sol","modelFamilyId":2},"permissionProfileId":"full"},"acceptedView":{"interactionNodeId":"node:1","rootAction":{"id":"action:1","clientKey":"action:1","sourceNodeId":"node:1","kind":"navigate","relation":"expand","label":"View compute-market analysis","variant":"pill","targetLayerId":"layer:1","state":"accepted"},"rootLayerId":"layer:1","layers":[{"layer":{"defaultNodeId":"node:2","id":"layer:1","clientKey":"layer:1","nodes":["node:2","node:3","node:4","node:5"],"edges":["edge:1","edge:2","edge:3"],"layout":{"version":1,"placements":[{"nodeId":"node:2","x":0.5,"y":0.14},{"nodeId":"node:3","x":0.16,"y":0.75},{"nodeId":"node:4","x":0.5,"y":0.9},{"nodeId":"node:5","x":0.84,"y":0.75}],"edgeShape":"arc-outward"},"state":"accepted"},"nodes":[{"id":"node:2","clientKey":"node:2","kind":"analysis","icon":"compass","title":"Bottom line","detail":"As of September 2026, compute markets are real but segmented. Frontier training is guaranteed infrastructure; flexible batch and inference are increasingly schedulable and brokered. Federation is likelier than one universal spot exchange.","authoredDetail":{"assets":[],"components":[{"css":"[data-relayer-theme=\"light\"]{display:block;--b:#f7f8f5;--s:#fff;--r:#edf1ec;--t:#172421;--m:#53635e;--o:#b8c5be;--a:#116a56;background-color:var(--b);color:var(--t)}[data-relayer-theme=\"dark\"]{display:block;--b:#111715;--s:#17201d;--r:#202c28;--t:#edf4f1;--m:#afbeb8;--o:#536860;--a:#74d3b7;background-color:var(--b);color:var(--t)}main,section,article,aside,nav,table{box-sizing:border-box}.p{display:grid;gap:.8rem;padding:.95rem;color:var(--t);background-color:var(--b)}.h{display:grid;gap:.6rem;padding:.85rem;border:1px solid var(--o);border-radius:.8rem;background-color:var(--s)}.e{margin:0;color:var(--a);font-size:.75rem;font-weight:700;letter-spacing:.08em;text-transform:uppercase}h2,p,ol{margin:0}h2{font-size:1.3rem}p,li,td,th{line-height:1.45}.g{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:.6rem}.c{display:grid;gap:.3rem;padding:.65rem;border:1px solid var(--o);border-radius:.65rem;background-color:var(--r)}.x{font-size:1.15rem;font-weight:750;color:var(--a)}.q{padding:.7rem;border-left:.25rem solid var(--a);background-color:var(--r)}table{width:100%;font-size:.85rem}th,td{padding:.42rem;border-bottom:1px solid var(--o);text-align:left;vertical-align:top}th,.m{color:var(--m)}nav{display:grid;gap:.45rem}button,a{color:var(--t);background-color:var(--r);border:1px solid var(--o);border-radius:.55rem;padding:.55rem .65rem;font:inherit;text-align:left;text-decoration:none}button:hover:enabled,a:hover{border-color:var(--a)}button:focus-visible,a:focus-visible{outline:2px solid var(--a);outline-offset:2px}@media (max-width:620px){.g{grid-template-columns:1fr}}","html":"

Snapshot · 30 Sep 2026

Compute trades—but not yet as one commodity

The frontier product is a reliable cluster. The emerging commodity is a portable workload.

Nowreservations · spot · schedulers · bilateral contracts

Limitpower · interconnect · software lock-in · financing

","id":"main","order":0}],"integritySha256":"0d65847e37d82018e8b6c329b2e1362d09e72ec5417f3f83cf47eb6a4f51d736","mounts":[{"capability":{"action":{"clientKey":"action:2","sourceLayer":{"clientKey":"layer:1"},"sourceNode":{"clientKey":"node:2"}},"kind":"reference"},"componentId":"main","host":"button","id":"m_b861293037e37734","kind":"capability"}],"version":1},"state":"accepted"},{"id":"node:3","clientKey":"node:3","kind":"analysis","icon":"layout-grid","title":"What trades today","detail":"Frontier training uses reserved clusters and take-or-pay; fine-tuning and batch use queued, calendar, and spot capacity; inference increasingly uses serverless APIs and multi-provider routing.","authoredDetail":{"assets":[],"components":[{"css":"[data-relayer-theme=\"light\"]{display:block;--b:#f7f8f5;--s:#fff;--r:#edf1ec;--t:#172421;--m:#53635e;--o:#b8c5be;--a:#116a56;background-color:var(--b);color:var(--t)}[data-relayer-theme=\"dark\"]{display:block;--b:#111715;--s:#17201d;--r:#202c28;--t:#edf4f1;--m:#afbeb8;--o:#536860;--a:#74d3b7;background-color:var(--b);color:var(--t)}main,section,article,aside,nav,table{box-sizing:border-box}.p{display:grid;gap:.8rem;padding:.95rem;color:var(--t);background-color:var(--b)}.h{display:grid;gap:.6rem;padding:.85rem;border:1px solid var(--o);border-radius:.8rem;background-color:var(--s)}.e{margin:0;color:var(--a);font-size:.75rem;font-weight:700;letter-spacing:.08em;text-transform:uppercase}h2,p,ol{margin:0}h2{font-size:1.3rem}p,li,td,th{line-height:1.45}.g{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:.6rem}.c{display:grid;gap:.3rem;padding:.65rem;border:1px solid var(--o);border-radius:.65rem;background-color:var(--r)}.x{font-size:1.15rem;font-weight:750;color:var(--a)}.q{padding:.7rem;border-left:.25rem solid var(--a);background-color:var(--r)}table{width:100%;font-size:.85rem}th,td{padding:.42rem;border-bottom:1px solid var(--o);text-align:left;vertical-align:top}th,.m{color:var(--m)}nav{display:grid;gap:.45rem}button,a{color:var(--t);background-color:var(--r);border:1px solid var(--o);border-radius:.55rem;padding:.55rem .65rem;font:inherit;text-align:left;text-decoration:none}button:hover:enabled,a:hover{border-color:var(--a)}button:focus-visible,a:focus-visible{outline:2px solid var(--a);outline-offset:2px}@media (max-width:620px){.g{grid-template-columns:1fr}}","html":"

Market maturity

Liquidity follows workload tolerance

Workload	Clearing model	Tradability
Frontier training	reserved / self-build	low
Fine-tuning and batch	calendar / queued / spot	medium
Inference	serverless / routers	medium → high
Rendering and tolerant jobs	open marketplaces	high

State of the art: portfolio procurement plus workload-aware scheduling, not one clearing price.

","id":"main","order":0}],"integritySha256":"780f00c40d62de2bc07867cebcb819170c5b9514c80d524cf313f42f4bd52efe","mounts":[],"version":1},"state":"accepted"},{"id":"node:4","clientKey":"node:4","kind":"analysis","icon":"users","title":"Who matters","detail":"Value is split among accelerator vendors, hyperscalers, AI-native clouds, brokers and marketplaces, decentralized networks, and the labs and enterprises anchoring demand.","authoredDetail":{"assets":[],"components":[{"css":"[data-relayer-theme=\"light\"]{display:block;--b:#f7f8f5;--s:#fff;--r:#edf1ec;--t:#172421;--m:#53635e;--o:#b8c5be;--a:#116a56;background-color:var(--b);color:var(--t)}[data-relayer-theme=\"dark\"]{display:block;--b:#111715;--s:#17201d;--r:#202c28;--t:#edf4f1;--m:#afbeb8;--o:#536860;--a:#74d3b7;background-color:var(--b);color:var(--t)}main,section,article,aside,nav,table{box-sizing:border-box}.p{display:grid;gap:.8rem;padding:.95rem;color:var(--t);background-color:var(--b)}.h{display:grid;gap:.6rem;padding:.85rem;border:1px solid var(--o);border-radius:.8rem;background-color:var(--s)}.e{margin:0;color:var(--a);font-size:.75rem;font-weight:700;letter-spacing:.08em;text-transform:uppercase}h2,p,ol{margin:0}h2{font-size:1.3rem}p,li,td,th{line-height:1.45}.g{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:.6rem}.c{display:grid;gap:.3rem;padding:.65rem;border:1px solid var(--o);border-radius:.65rem;background-color:var(--r)}.x{font-size:1.15rem;font-weight:750;color:var(--a)}.q{padding:.7rem;border-left:.25rem solid var(--a);background-color:var(--r)}table{width:100%;font-size:.85rem}th,td{padding:.42rem;border-bottom:1px solid var(--o);text-align:left;vertical-align:top}th,.m{color:var(--m)}nav{display:grid;gap:.45rem}button,a{color:var(--t);background-color:var(--r);border:1px solid var(--o);border-radius:.55rem;padding:.55rem .65rem;font:inherit;text-align:left;text-decoration:none}button:hover:enabled,a:hover{border-color:var(--a)}button:focus-visible,a:focus-visible{outline:2px solid var(--a);outline-offset:2px}@media (max-width:620px){.g{grid-template-columns:1fr}}","html":"

Competitive map

Five groups shape price

SiliconNVIDIA · AMD · custom accelerators

HyperscalersAWS · Azure · Google · Oracle

AI-native cloudsCoreWeave · Crusoe · Lambda · Nebius · Nscale

InterfacesLepton · brokers · exchanges · open networks

","id":"main","order":0}],"integritySha256":"01318e2dd7cb3ea5b16927dc86854af0c50f902a69a85f7f2c6819345305e3fb","mounts":[{"capability":{"action":{"clientKey":"action:3","sourceLayer":{"clientKey":"layer:1"},"sourceNode":{"clientKey":"node:4"}},"kind":"expand"},"componentId":"main","host":"button","id":"m_b1789e5a6479751d","kind":"capability"}],"version":1},"state":"accepted"},{"id":"node:5","clientKey":"node:5","kind":"analysis","icon":"bolt","title":"Expert outlook","detail":"IEA and Epoch AI imply rapid physical scaling with power as the gating variable. Expect reservations for scarce clusters, dynamic routing for inference, and regional premiums for energy and sovereignty.","authoredDetail":{"assets":[],"components":[{"css":"[data-relayer-theme=\"light\"]{display:block;--b:#f7f8f5;--s:#fff;--r:#edf1ec;--t:#172421;--m:#53635e;--o:#b8c5be;--a:#116a56;background-color:var(--b);color:var(--t)}[data-relayer-theme=\"dark\"]{display:block;--b:#111715;--s:#17201d;--r:#202c28;--t:#edf4f1;--m:#afbeb8;--o:#536860;--a:#74d3b7;background-color:var(--b);color:var(--t)}main,section,article,aside,nav,table{box-sizing:border-box}.p{display:grid;gap:.8rem;padding:.95rem;color:var(--t);background-color:var(--b)}.h{display:grid;gap:.6rem;padding:.85rem;border:1px solid var(--o);border-radius:.8rem;background-color:var(--s)}.e{margin:0;color:var(--a);font-size:.75rem;font-weight:700;letter-spacing:.08em;text-transform:uppercase}h2,p,ol{margin:0}h2{font-size:1.3rem}p,li,td,th{line-height:1.45}.g{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:.6rem}.c{display:grid;gap:.3rem;padding:.65rem;border:1px solid var(--o);border-radius:.65rem;background-color:var(--r)}.x{font-size:1.15rem;font-weight:750;color:var(--a)}.q{padding:.7rem;border-left:.25rem solid var(--a);background-color:var(--r)}table{width:100%;font-size:.85rem}th,td{padding:.42rem;border-bottom:1px solid var(--o);text-align:left;vertical-align:top}th,.m{color:var(--m)}nav{display:grid;gap:.45rem}button,a{color:var(--t);background-color:var(--r);border:1px solid var(--o);border-radius:.55rem;padding:.55rem .65rem;font:inherit;text-align:left;text-decoration:none}button:hover:enabled,a:hover{border-color:var(--a)}button:focus-visible,a:focus-visible{outline:2px solid var(--a);outline-offset:2px}@media (max-width:620px){.g{grid-template-columns:1fr}}","html":"

Expert-informed outlook

Power becomes a market variable

2×IEA: total data-center electricity, 2025–2030

3×IEA: AI-focused data-center electricity

4–16 GWEpoch: possible largest 2030 training run

Structurereservations for clusters; dynamic routing for inference

","id":"main","order":0}],"integritySha256":"829e80b1bc428daa072d68c6206ec9d9572d5570f93519e232692ff70cf70178","mounts":[{"capability":{"action":{"clientKey":"action:4","sourceLayer":{"clientKey":"layer:1"},"sourceNode":{"clientKey":"node:5"}},"kind":"expand"},"componentId":"main","host":"button","id":"m_a3ba1f0111be30d1","kind":"capability"}],"version":1},"state":"accepted"}],"edges":[{"id":"edge:1","endpoints":["node:2","node:3"],"state":"accepted"},{"id":"edge:2","endpoints":["node:2","node:4"],"state":"accepted"},{"id":"edge:3","endpoints":["node:2","node:5"],"state":"accepted"}],"actions":[{"id":"action:2","clientKey":"action:2","sourceNodeId":"node:2","sourceLayerId":"layer:1","kind":"navigate","relation":"reference","label":"Open evidence and sources","variant":"pill","targetLayerId":"layer:2","state":"accepted"},{"id":"action:3","clientKey":"action:3","sourceNodeId":"node:4","sourceLayerId":"layer:1","kind":"navigate","relation":"expand","label":"Explore the player map","variant":"pill","targetLayerId":"layer:3","state":"accepted"},{"id":"action:4","clientKey":"action:4","sourceNodeId":"node:5","sourceLayerId":"layer:1","kind":"navigate","relation":"expand","label":"Explore 2030 scenarios","variant":"pill","targetLayerId":"layer:4","state":"accepted"}]},{"layer":{"defaultNodeId":"node:6","id":"layer:2","clientKey":"layer:2","nodes":["node:6"],"edges":[],"layout":{"version":1,"placements":[{"nodeId":"node:6","x":0.5,"y":0.5}],"edgeShape":"default"},"state":"accepted"},"nodes":[{"id":"node:6","clientKey":"node:6","kind":"evidence","icon":"book-open-text","title":"Evidence & sources","detail":"IEA, Epoch AI, Stanford HAI, provider product documentation, and SEC filings support the market map. Forecasts remain uncertain, especially where power feasibility and efficiency interact.","authoredDetail":{"assets":[],"components":[{"css":"[data-relayer-theme=\"light\"]{display:block;--b:#f7f8f5;--s:#fff;--r:#edf1ec;--t:#172421;--m:#53635e;--o:#b8c5be;--a:#116a56;background-color:var(--b);color:var(--t)}[data-relayer-theme=\"dark\"]{display:block;--b:#111715;--s:#17201d;--r:#202c28;--t:#edf4f1;--m:#afbeb8;--o:#536860;--a:#74d3b7;background-color:var(--b);color:var(--t)}main,section,article,aside,nav,table{box-sizing:border-box}.p{display:grid;gap:.8rem;padding:.95rem;color:var(--t);background-color:var(--b)}.h{display:grid;gap:.6rem;padding:.85rem;border:1px solid var(--o);border-radius:.8rem;background-color:var(--s)}.e{margin:0;color:var(--a);font-size:.75rem;font-weight:700;letter-spacing:.08em;text-transform:uppercase}h2,p,ol{margin:0}h2{font-size:1.3rem}p,li,td,th{line-height:1.45}.g{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:.6rem}.c{display:grid;gap:.3rem;padding:.65rem;border:1px solid var(--o);border-radius:.65rem;background-color:var(--r)}.x{font-size:1.15rem;font-weight:750;color:var(--a)}.q{padding:.7rem;border-left:.25rem solid var(--a);background-color:var(--r)}table{width:100%;font-size:.85rem}th,td{padding:.42rem;border-bottom:1px solid var(--o);text-align:left;vertical-align:top}th,.m{color:var(--m)}nav{display:grid;gap:.45rem}button,a{color:var(--t);background-color:var(--r);border:1px solid var(--o);border-radius:.55rem;padding:.55rem .65rem;font:inherit;text-align:left;text-decoration:none}button:hover:enabled,a:hover{border-color:var(--a)}button:focus-visible,a:focus-visible{outline:2px solid var(--a);outline-offset:2px}@media (max-width:620px){.g{grid-template-columns:1fr}}","html":"

Observed evidence

Products are real; the endpoint is uncertain

~950 TWhIEA 2030 data-center electricity base case

4–16 GWEpoch range for one largest 2030 training run

AWS Capacity Blocks, Google DWS, and NVIDIA Lepton show reservation, scheduling, and federation already shipping. CoreWeave’s filing shows take-or-pay financing and concentration risk.

","id":"main","order":0}],"integritySha256":"2af4eec34a7ce640a3c4d78a142c17c17edb27b0dd133d86be1fc7e8c982cbb0","mounts":[{"capability":{"href":"https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary","kind":"link"},"componentId":"main","host":"a","id":"m_89df27ae4d581590","kind":"capability"},{"capability":{"href":"https://epoch.ai/publications/power-demands-of-frontier-ai-training","kind":"link"},"componentId":"main","host":"a","id":"m_217f9852cbf109e7","kind":"capability"},{"capability":{"href":"https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf","kind":"link"},"componentId":"main","host":"a","id":"m_9e9f9d10cab677ad","kind":"capability"},{"capability":{"href":"https://aws.amazon.com/ec2/capacityblocks/","kind":"link"},"componentId":"main","host":"a","id":"m_161b21b5199b12a5","kind":"capability"},{"capability":{"href":"https://cloud.google.com/products/dws/pricing","kind":"link"},"componentId":"main","host":"a","id":"m_beae92102853cbdd","kind":"capability"},{"capability":{"href":"https://nvidianews.nvidia.com/news/nvidia-announces-dgx-cloud-lepton-to-connect-developers-to-nvidias-global-compute-ecosystem","kind":"link"},"componentId":"main","host":"a","id":"m_5c9a4ab0656f545b","kind":"capability"},{"capability":{"href":"https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm","kind":"link"},"componentId":"main","host":"a","id":"m_dc0535f8dce5f29b","kind":"capability"}],"version":1},"state":"accepted"}],"edges":[],"actions":[]},{"layer":{"defaultNodeId":"node:7","id":"layer:3","clientKey":"layer:3","nodes":["node:7","node:8","node:9","node:10","node:11"],"edges":["edge:4","edge:5","edge:6","edge:7"],"layout":{"version":1,"placements":[{"nodeId":"node:7","x":0.5,"y":0.12},{"nodeId":"node:8","x":0.15,"y":0.58},{"nodeId":"node:9","x":0.35,"y":0.88},{"nodeId":"node:10","x":0.68,"y":0.86},{"nodeId":"node:11","x":0.88,"y":0.52}],"edgeShape":"arc-outward"},"state":"accepted"},"nodes":[{"id":"node:7","clientKey":"node:7","kind":"analysis","icon":"layers","title":"Compute value chain","detail":"Compute is a stack: silicon and power, cluster operators, market interfaces, and workload buyers. A usable cluster—not a bare GPU—is the scarce product.","authoredDetail":{"assets":[],"components":[{"css":"[data-relayer-theme=\"light\"]{display:block;--b:#f7f8f5;--s:#fff;--r:#edf1ec;--t:#172421;--m:#53635e;--o:#b8c5be;--a:#116a56;background-color:var(--b);color:var(--t)}[data-relayer-theme=\"dark\"]{display:block;--b:#111715;--s:#17201d;--r:#202c28;--t:#edf4f1;--m:#afbeb8;--o:#536860;--a:#74d3b7;background-color:var(--b);color:var(--t)}main,section,article,aside,nav,table{box-sizing:border-box}.p{display:grid;gap:.8rem;padding:.95rem;color:var(--t);background-color:var(--b)}.h{display:grid;gap:.6rem;padding:.85rem;border:1px solid var(--o);border-radius:.8rem;background-color:var(--s)}.e{margin:0;color:var(--a);font-size:.75rem;font-weight:700;letter-spacing:.08em;text-transform:uppercase}h2,p,ul,ol{margin:0}h2{font-size:1.3rem}p,li,td,th{line-height:1.45}.g{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:.6rem}.c{display:grid;gap:.3rem;padding:.65rem;border:1px solid var(--o);border-radius:.65rem;background-color:var(--r)}.x{font-size:1.15rem;font-weight:750;color:var(--a)}.q{padding:.7rem;border-left:.25rem solid var(--a);background-color:var(--r)}table{width:100%;font-size:.85rem}th,td{padding:.42rem;border-bottom:1px solid var(--o);text-align:left;vertical-align:top}th,.m{color:var(--m)}nav{display:grid;gap:.45rem}button,a{color:var(--t);background-color:var(--r);border:1px solid var(--o);border-radius:.55rem;padding:.55rem .65rem;font:inherit;text-align:left;text-decoration:none}button:hover:enabled,a:hover{border-color:var(--a)}button:focus-visible,a:focus-visible{outline:2px solid var(--a);outline-offset:2px}@media (max-width:620px){.g{grid-template-columns:1fr}}","html":"

Value chain

The scarce unit is a usable cluster

A GPU-hour is not fungible until buyers can compare fabric, memory, storage, software, security, location, uptime, and start-time guarantees.

Upstreamchips · memory · networking · power

Interfacescheduling · brokerage · APIs

","id":"main","order":0}],"integritySha256":"3a559b8b8df702bb898c7bcd6f387824ba24cb0c1b48e46a9f071e999893025e","mounts":[],"version":1},"state":"accepted"},{"id":"node:8","clientKey":"node:8","kind":"analysis","icon":"cpu","title":"Silicon & systems","detail":"NVIDIA remains the compatibility anchor. AMD Instinct, Google TPU, AWS Trainium, and other custom accelerators are the principal alternatives; networking, memory, cooling, and power are inseparable.","authoredDetail":{"assets":[],"components":[{"css":"[data-relayer-theme=\"light\"]{display:block;--b:#f7f8f5;--s:#fff;--r:#edf1ec;--t:#172421;--m:#53635e;--o:#b8c5be;--a:#116a56;background-color:var(--b);color:var(--t)}[data-relayer-theme=\"dark\"]{display:block;--b:#111715;--s:#17201d;--r:#202c28;--t:#edf4f1;--m:#afbeb8;--o:#536860;--a:#74d3b7;background-color:var(--b);color:var(--t)}main,section,article,aside,nav,table{box-sizing:border-box}.p{display:grid;gap:.8rem;padding:.95rem;color:var(--t);background-color:var(--b)}.h{display:grid;gap:.6rem;padding:.85rem;border:1px solid var(--o);border-radius:.8rem;background-color:var(--s)}.e{margin:0;color:var(--a);font-size:.75rem;font-weight:700;letter-spacing:.08em;text-transform:uppercase}h2,p,ul,ol{margin:0}h2{font-size:1.3rem}p,li,td,th{line-height:1.45}.g{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:.6rem}.c{display:grid;gap:.3rem;padding:.65rem;border:1px solid var(--o);border-radius:.65rem;background-color:var(--r)}.x{font-size:1.15rem;font-weight:750;color:var(--a)}.q{padding:.7rem;border-left:.25rem solid var(--a);background-color:var(--r)}table{width:100%;font-size:.85rem}th,td{padding:.42rem;border-bottom:1px solid var(--o);text-align:left;vertical-align:top}th,.m{color:var(--m)}nav{display:grid;gap:.45rem}button,a{color:var(--t);background-color:var(--r);border:1px solid var(--o);border-radius:.55rem;padding:.55rem .65rem;font:inherit;text-align:left;text-decoration:none}button:hover:enabled,a:hover{border-color:var(--a)}button:focus-visible,a:focus-visible{outline:2px solid var(--a);outline-offset:2px}@media (max-width:620px){.g{grid-template-columns:1fr}}","html":"

Supply

Accelerators diversify slowly

NVIDIACUDA and rack-scale systems

AlternativesAMD Instinct · Google TPU · AWS Trainium · Microsoft Maia

Delivered goodput and porting cost matter more than peak FLOPS.

","id":"main","order":0}],"integritySha256":"d6a90e1fa29b0fc4f789cd64de5812b512d0c8adf39ce49244100ad3f961a066","mounts":[],"version":1},"state":"accepted"},{"id":"node:9","clientKey":"node:9","kind":"analysis","icon":"cloud","title":"Cloud operators","detail":"Hyperscalers AWS, Azure, Google Cloud, and Oracle coexist with AI-native clouds such as CoreWeave, Crusoe, Lambda, Nebius, Nscale, Fluidstack, and IREN.","authoredDetail":{"assets":[],"components":[{"css":"[data-relayer-theme=\"light\"]{display:block;--b:#f7f8f5;--s:#fff;--r:#edf1ec;--t:#172421;--m:#53635e;--o:#b8c5be;--a:#116a56;background-color:var(--b);color:var(--t)}[data-relayer-theme=\"dark\"]{display:block;--b:#111715;--s:#17201d;--r:#202c28;--t:#edf4f1;--m:#afbeb8;--o:#536860;--a:#74d3b7;background-color:var(--b);color:var(--t)}main,section,article,aside,nav,table{box-sizing:border-box}.p{display:grid;gap:.8rem;padding:.95rem;color:var(--t);background-color:var(--b)}.h{display:grid;gap:.6rem;padding:.85rem;border:1px solid var(--o);border-radius:.8rem;background-color:var(--s)}.e{margin:0;color:var(--a);font-size:.75rem;font-weight:700;letter-spacing:.08em;text-transform:uppercase}h2,p,ul,ol{margin:0}h2{font-size:1.3rem}p,li,td,th{line-height:1.45}.g{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:.6rem}.c{display:grid;gap:.3rem;padding:.65rem;border:1px solid var(--o);border-radius:.65rem;background-color:var(--r)}.x{font-size:1.15rem;font-weight:750;color:var(--a)}.q{padding:.7rem;border-left:.25rem solid var(--a);background-color:var(--r)}table{width:100%;font-size:.85rem}th,td{padding:.42rem;border-bottom:1px solid var(--o);text-align:left;vertical-align:top}th,.m{color:var(--m)}nav{display:grid;gap:.45rem}button,a{color:var(--t);background-color:var(--r);border:1px solid var(--o);border-radius:.55rem;padding:.55rem .65rem;font:inherit;text-align:left;text-decoration:none}button:hover:enabled,a:hover{border-color:var(--a)}button:focus-visible,a:focus-visible{outline:2px solid var(--a);outline-offset:2px}@media (max-width:620px){.g{grid-template-columns:1fr}}","html":"

Operators

Two cloud models coexist

HyperscalersAWS · Azure · Google Cloud · OracleBroad services and enterprise trust.

AI-native cloudsCoreWeave · Crusoe · Lambda · Nebius · Nscale · Fluidstack · IRENFast deployment; more capital risk.

","id":"main","order":0}],"integritySha256":"4909d49ac1532dd78d8ce1bc224b5ce5ddc1b3b07da143a6952e72e34d4410a8","mounts":[],"version":1},"state":"accepted"},{"id":"node:10","clientKey":"node:10","kind":"analysis","icon":"network","title":"Markets & brokers","detail":"AWS Spot and Capacity Blocks, Google DWS, NVIDIA DGX Cloud Lepton, Compute Exchange, Vast.ai, Runpod, and other brokers provide increasingly market-like access.","authoredDetail":{"assets":[],"components":[{"css":"[data-relayer-theme=\"light\"]{display:block;--b:#f7f8f5;--s:#fff;--r:#edf1ec;--t:#172421;--m:#53635e;--o:#b8c5be;--a:#116a56;background-color:var(--b);color:var(--t)}[data-relayer-theme=\"dark\"]{display:block;--b:#111715;--s:#17201d;--r:#202c28;--t:#edf4f1;--m:#afbeb8;--o:#536860;--a:#74d3b7;background-color:var(--b);color:var(--t)}main,section,article,aside,nav,table{box-sizing:border-box}.p{display:grid;gap:.8rem;padding:.95rem;color:var(--t);background-color:var(--b)}.h{display:grid;gap:.6rem;padding:.85rem;border:1px solid var(--o);border-radius:.8rem;background-color:var(--s)}.e{margin:0;color:var(--a);font-size:.75rem;font-weight:700;letter-spacing:.08em;text-transform:uppercase}h2,p,ul,ol{margin:0}h2{font-size:1.3rem}p,li,td,th{line-height:1.45}.g{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:.6rem}.c{display:grid;gap:.3rem;padding:.65rem;border:1px solid var(--o);border-radius:.65rem;background-color:var(--r)}.x{font-size:1.15rem;font-weight:750;color:var(--a)}.q{padding:.7rem;border-left:.25rem solid var(--a);background-color:var(--r)}table{width:100%;font-size:.85rem}th,td{padding:.42rem;border-bottom:1px solid var(--o);text-align:left;vertical-align:top}th,.m{color:var(--m)}nav{display:grid;gap:.45rem}button,a{color:var(--t);background-color:var(--r);border:1px solid var(--o);border-radius:.55rem;padding:.55rem .65rem;font:inherit;text-align:left;text-decoration:none}button:hover:enabled,a:hover{border-color:var(--a)}button:focus-visible,a:focus-visible{outline:2px solid var(--a);outline-offset:2px}@media (max-width:620px){.g{grid-template-columns:1fr}}","html":"

Market interfaces

Scheduling is ahead of exchange trading

Provider-native	AWS Spot / Blocks · Google DWS
Federated	NVIDIA Lepton
Open	Vast.ai · Compute Exchange
Developer	Runpod and peers

The frontier is quality-adjusted procurement, not merely $/GPU-hour.

","id":"main","order":0}],"integritySha256":"b2cc3f01d48bc71c9f6079e77b0e435d5e1f18e09670bbdd258397d64e9ea967","mounts":[],"version":1},"state":"accepted"},{"id":"node:11","clientKey":"node:11","kind":"analysis","icon":"globe","title":"Open networks","detail":"Akash, io.net, Golem, Render Network, and similar systems aggregate independent supply. Their near-term fit is tolerant batch, rendering, and inference—not tightly synchronized frontier training.","authoredDetail":{"assets":[],"components":[{"css":"[data-relayer-theme=\"light\"]{display:block;--b:#f7f8f5;--s:#fff;--r:#edf1ec;--t:#172421;--m:#53635e;--o:#b8c5be;--a:#116a56;background-color:var(--b);color:var(--t)}[data-relayer-theme=\"dark\"]{display:block;--b:#111715;--s:#17201d;--r:#202c28;--t:#edf4f1;--m:#afbeb8;--o:#536860;--a:#74d3b7;background-color:var(--b);color:var(--t)}main,section,article,aside,nav,table{box-sizing:border-box}.p{display:grid;gap:.8rem;padding:.95rem;color:var(--t);background-color:var(--b)}.h{display:grid;gap:.6rem;padding:.85rem;border:1px solid var(--o);border-radius:.8rem;background-color:var(--s)}.e{margin:0;color:var(--a);font-size:.75rem;font-weight:700;letter-spacing:.08em;text-transform:uppercase}h2,p,ul,ol{margin:0}h2{font-size:1.3rem}p,li,td,th{line-height:1.45}.g{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:.6rem}.c{display:grid;gap:.3rem;padding:.65rem;border:1px solid var(--o);border-radius:.65rem;background-color:var(--r)}.x{font-size:1.15rem;font-weight:750;color:var(--a)}.q{padding:.7rem;border-left:.25rem solid var(--a);background-color:var(--r)}table{width:100%;font-size:.85rem}th,td{padding:.42rem;border-bottom:1px solid var(--o);text-align:left;vertical-align:top}th,.m{color:var(--m)}nav{display:grid;gap:.45rem}button,a{color:var(--t);background-color:var(--r);border:1px solid var(--o);border-radius:.55rem;padding:.55rem .65rem;font:inherit;text-align:left;text-decoration:none}button:hover:enabled,a:hover{border-color:var(--a)}button:focus-visible,a:focus-visible{outline:2px solid var(--a);outline-offset:2px}@media (max-width:620px){.g{grid-template-columns:1fr}}","html":"

Long-tail supply

Decentralized markets are workload-selective

Best fitrendering · parallel batch · burst inference

Constraintstrust · egress · locality · uptime · interconnect

","id":"main","order":0}],"integritySha256":"7fcaf68f19593c3c548fd0195f3d62940d8e4955a99dcdf2c95dd8c062c905ff","mounts":[],"version":1},"state":"accepted"}],"edges":[{"id":"edge:4","endpoints":["node:7","node:8"],"state":"accepted"},{"id":"edge:5","endpoints":["node:7","node:9"],"state":"accepted"},{"id":"edge:6","endpoints":["node:7","node:10"],"state":"accepted"},{"id":"edge:7","endpoints":["node:7","node:11"],"state":"accepted"}],"actions":[]},{"layer":{"defaultNodeId":"node:12","id":"layer:4","clientKey":"layer:4","nodes":["node:12"],"edges":[],"layout":{"version":1,"placements":[{"nodeId":"node:12","x":0.5,"y":0.5}],"edgeShape":"default"},"state":"accepted"},"nodes":[{"id":"node:12","clientKey":"node:12","kind":"analysis","icon":"route","title":"2030 scenarios","detail":"The base case is federation: reserved frontier clusters, scheduler-cleared flexible jobs, and outcome-priced inference. Faster portability could create liquid capacity markets; power and policy constraints could instead deepen regional fragmentation.","authoredDetail":{"assets":[],"components":[{"css":"[data-relayer-theme=\"light\"]{display:block;--b:#f7f8f5;--s:#fff;--r:#edf1ec;--t:#172421;--m:#53635e;--o:#b8c5be;--a:#116a56;background-color:var(--b);color:var(--t)}[data-relayer-theme=\"dark\"]{display:block;--b:#111715;--s:#17201d;--r:#202c28;--t:#edf4f1;--m:#afbeb8;--o:#536860;--a:#74d3b7;background-color:var(--b);color:var(--t)}main,section,article,aside,nav,table{box-sizing:border-box}.p{display:grid;gap:.8rem;padding:.95rem;color:var(--t);background-color:var(--b)}.h{display:grid;gap:.6rem;padding:.85rem;border:1px solid var(--o);border-radius:.8rem;background-color:var(--s)}.e{margin:0;color:var(--a);font-size:.75rem;font-weight:700;letter-spacing:.08em;text-transform:uppercase}h2,p,ol{margin:0}h2{font-size:1.3rem}p,li,td,th{line-height:1.45}.g{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:.6rem}.c{display:grid;gap:.3rem;padding:.65rem;border:1px solid var(--o);border-radius:.65rem;background-color:var(--r)}.x{font-size:1.15rem;font-weight:750;color:var(--a)}.q{padding:.7rem;border-left:.25rem solid var(--a);background-color:var(--r)}table{width:100%;font-size:.85rem}th,td{padding:.42rem;border-bottom:1px solid var(--o);text-align:left;vertical-align:top}th,.m{color:var(--m)}nav{display:grid;gap:.45rem}button,a{color:var(--t);background-color:var(--r);border:1px solid var(--o);border-radius:.55rem;padding:.55rem .65rem;font:inherit;text-align:left;text-decoration:none}button:hover:enabled,a:hover{border-color:var(--a)}button:focus-visible,a:focus-visible{outline:2px solid var(--a);outline-offset:2px}@media (max-width:620px){.g{grid-template-columns:1fr}}","html":"

Expert-informed scenarios

Federation is the center path

Base caselarge operators remain central; brokers make workloads more portable

Liquid upsideprices clear on tokens, goodput, latency, region, and carbon

Constraint casegrid queues, financing, export controls, and sovereignty regionalize capacity

Watchportability, utilization, contract duration, power lead times

","id":"main","order":0}],"integritySha256":"142975d16c935ef897c9876646d95ab4e1c6f7a42c51a4e1dbf3a45e5ca83b3e","mounts":[],"version":1},"state":"accepted"}],"edges":[],"actions":[]}]}}

This shared thread couldn’t be loaded