Startup and adjacent-provider landscape

As of: 2026-08-26. This is a taxonomy and watchlist, not an endorsement, market-share claim, or proof that every company remains independent. Company state changes quickly; material claims belong in index.json with a dated source.

Layer Representative companies/projects to monitor Expectation exposed
Neocloud / GPU capacity CoreWeave, Nebius, Crusoe, Lambda, Together AI, Voltage Park, Fluidstack, Applied Digital, Northern Data/Taiga, Nscale Capital-intensive capacity, customer concentration, financing, power access, cluster goodput, and managed Kubernetes/bare metal matter together.
Specialized inference silicon/cloud Groq, Cerebras, SambaNova, d-Matrix, Etched, Tenstorrent, Positron, FuriosaAI, Rebellions, SiMa.ai Low latency, deterministic execution, memory bandwidth, and phase specialization compete with general GPU ecosystem breadth.
Serving/runtime/orchestration vLLM, SGLang, TensorRT-LLM, NVIDIA Dynamo, llm-d, Ray Serve/Anyscale, BentoML, Baseten, Modal, Fireworks AI, Replicate, OctoAI lineage Continuous batching, KV state, disaggregation, routing, autoscaling, multi-tenancy, and observability are becoming a control plane.
Model/API routing and gateways OpenRouter, Portkey, Martian, Not Diamond, Cloudflare AI Gateway, Kong/Envoy ecosystem Providers, models, regions, cache state, price, quality, and policy vary per request; fallback can damage cache locality and provenance.
KV cache / memory / storage LMCache, Mooncake, DeepSeek 3FS, Alibaba Tair KVCache, Alluxio, Weka, VAST Data, Hammerspace Inference state is becoming a distributed storage/network problem, not only local HBM.
Networking/interconnect Broadcom, Marvell, Astera Labs, Enfabrica, Lightmatter, Celestial AI, Ayar Labs, Cornelis Networks, Xsight Labs Scale-up/scale-out fabrics, optics, congestion, collectives, and composability can gate useful accelerator output.
HBM, packaging, chiplets SK hynix, Samsung, Micron, TSMC/CoWoS, ASE, Amkor, Eliyan, UCIe ecosystem Memory and advanced packaging lead times can bind before nominal compute supply.
Optical components/CPO Coherent, Lumentum, Innolight, Fabrinet, POET, DustPhotonics, Ranovus Lasers, fiber, transceivers, substrates, and co-packaged optics become rack/fleet delivery constraints.
Cooling / heat reuse Vertiv, Schneider Electric, CoolIT, Submer, LiquidStack, Iceotope, ZutaCore, JetCool, Accelsius Rack density shifts cooling from facility overhead to platform design; water and heat rejection alter site feasibility.
Power generation / grid / storage Oklo, Kairos, X-energy, TerraPower, Last Energy, Bloom Energy, Crusoe, Fervo, Form Energy, Redwood Materials Firm generation, turbines, nuclear/geothermal timelines, batteries, and behind-the-meter systems determine delivery dates and social license.
Datacenter development / modular build Vantage, QTS, Digital Realty, Equinix, Aligned, DataBank, STACK, EdgeConneX, Crusoe, Lancium, Tract, PowerHouse, T5 Land, substations, transmission, construction labor, modules, financing, permitting, and tenant contracts form one pipeline.
Sovereign AI / regional clouds G42/Core42, TII, HUMAIN, Nscale, Scaleway, OVHcloud, Jio, Yotta, sovereign programs in Europe/Asia/Middle East Jurisdiction, residency, national capacity, export controls, and local language/product needs constrain placement.
Datacenter observability / digital twins NVIDIA DSX/Omniverse, Cadence, Siemens, Schneider, Vertiv, specialized DCIM vendors Power, thermal, network, workload, and maintenance policy increasingly need joint simulation and read-back.

What to record for each company or project

Realized startup/provider failures (#9363)

Entity Outcome Boundary
Untether AI Existing bankruptcy record: nearly $25M assets, >$128M unsecured liabilities, $103.6M deficiency Insolvency is not a measured amount of removed live accelerator capacity.
Builder.ai Insolvency proceedings after a cash crunch Company continuity does not quantify GPU fleet, customer workload loss, or final liquidation.
Equinix Metal Bare-metal service wind-down and $233M impairment disclosed for 2024 Product/service withdrawal is not closure of Equinix colocation sites.

Failure modes in startup coverage

  1. Treating funding or valuation as technical/product evidence.
  2. Counting announced GPUs/MW as online healthy capacity.
  3. Repeating vendor benchmark maxima without workload and quality constraints.
  4. Missing acquisitions, licensing deals, shutdowns, or customer-concentration risk.
  5. Calling every GPU lessor a differentiated cloud or every chip startup deployed.
  6. Treating multiple articles copied from one leak as independent rumor corroboration.
  7. Ignoring private debt, take-or-pay contracts, and power-delivery risk.

Lifecycle outcomes beyond funding and launches

Company / asset Initial category Observed outcome What survives What failed or changed
Untether AI In-memory/RISC-V AI accelerator Shutdown, engineering-team transfer to AMD, later bankruptcy Engineering talent may continue inside AMD. Standalone company and product support ended; reported liabilities exceeded assets.
Replicate Hosted model execution/API Joined Cloudflare Brand and API were promised to continue; model execution primitives gain network/platform integration. Independent ownership; transaction economics and long-term product boundary are undisclosed.
SchedMD / Slurm Cluster scheduling/workload management Acquired by NVIDIA Slurm is promised to remain open source and vendor-neutral. Independent vendor governance; neutrality must be observed, not assumed.
Groq IP/team Inference accelerator/cloud Non-exclusive NVIDIA license plus reported key-talent transfer Groq remained operational; licensing was not described as a full acquisition. Independent roadmap, staffing, and competitive boundary may change.

Failure and acquisition extraction contract

For each future event, record separately:

legal entity / team / IP / product / customer contracts / support / brand
announced / signed / closed / integrated / discontinued / bankrupt
transaction value / assets / liabilities / runway if public
customer migration and compatibility path
open-source governance and hardware neutrality
later outcome check

Do not label an acquihire, IP license, minority investment, product acquisition, or bankruptcy as a full-company acquisition unless the source says so.

Alternative-infrastructure census additions

Company Category Evidence state Quantified evidence Main denominator risk
Lambda Neocloud / AI factory Multi-year Microsoft contract announced; fleet financing expanded Tens of thousands of NVIDIA GPUs including GB300 NVL72; >$1.5B Series E and later $1B credit facility in adjacent disclosures Contract, financing, target GW, installed GPUs, accepted service, and useful goodput differ.
Applied Digital Power-first AI datacenter developer/operator Polaris Forge 1 Building 2 Phase 1 reported Ready for Service 75 MW new RFS; 175 MW campus live; 400 MW contracted full build Live critical IT MW does not reveal installed customer compute, PUE, utilization, or goodput.
Lightmatter Photonic scale-up interconnect M1000 described as production-ready/validated hardware 4,000 mm² die complex; 34 chiplets; 1,024 SerDes lanes; 256 fibers; up to 114 Tbps Peak component bandwidth and production-ready labeling do not prove shipments or application goodput.
Celestial AI Photonic fabric startup Acquired by Marvell; transaction closed February 2026 $1B cash-balance reduction; ~27M new diluted shares; projected $500M annualized run rate in FY2028 Q4 and $1B in FY2029 Q4 Purchase price and earnout are not product shipment, qualification, or technical validation.
Fireworks AI Inference platform Large financing plus company-reported production scale $1.505B Series D; $17.5B valuation; >$1B annualized run rate; >40T tokens/day; >95% customer-specialized Company token/revenue definitions, customer concentration, physical fleet, cache, quality, margin, and profitability are undisclosed.

Category-specific checks