AI-crypto has found its cleanest sales deck: decentralized compute, decentralized inference, decentralized intelligence. Akash calls itself a decentralized cloud marketplace, Render describes a peer-to-peer GPU marketplace, Aethir markets distributed enterprise GPU cloud infrastructure, and Bittensor frames itself as an open platform for digital commodities including inference and compute.
The ambition is legitimate. The branding, however, is doing too much work. Decentralized AI increasingly depends on centralized-grade machinery, because serious model workloads require premium chips, industrial facilities, power contracts, bandwidth, cooling, and operational reliability that hobbyist networks cannot simply wish into existence for mainstream enterprise demand.
That physical constraint matters because AI is not secured by vibes or token incentives. The OECD notes that GPUs remain the most used chips for AI tasks in data centers and that Nvidia has been estimated at over 80% share for AI GPU chips, while the largest three cloud providers hold over 60% of global cloud market share. McKinsey projects AI data-center capex needs of $5.2 trillion by 2030. The supply chain is already oligopolistic, so crypto protocols that buy, rent, route, or tokenize scarce compute are entering a market shaped by capital intensity, not decentralization purity.
The contradiction becomes clearest when projects boast about scale. Aethir says it supports more than 440,000 high-performance GPU containers across 200 locations in 94 countries, including thousands of Nvidia H100, H200, B200, and B300 units. That sounds distributed, and in a geographic sense it may be. But enterprise-grade AI capacity is still clustered around professional hosts, procurement relationships, and data-center economics. A network can be globally distributed yet economically concentrated, which means the user sees a tokenized interface while the underlying leverage remains with whoever controls the racks, chips, and uptime.
The Token Does Not Decentralize the Rack
Blockchains decentralize consensus by making validation relatively legible. AI compute is messier. Training and inference require latency management, memory bandwidth, model checkpoints, specialized software stacks, data security, and predictable throughput. Bittensor’s documentation says subnets use miners to produce commodities and validators to evaluate their work, which is an elegant market design. Still, evaluation does not erase infrastructure dependency. AI workloads reward the biggest operators first, because the best hardware, lowest latency, and highest reliability usually sit with entities already capable of financing serious GPU footprints before rewards arrive, and staying online during demand spikes.
This is why the question, “Are these just cloud providers with tokens?” is uncomfortable but necessary. The answer is not entirely yes. Open marketplaces can improve price discovery, reduce platform lock-in, and let smaller buyers access compute without negotiating directly with hyperscalers. That is useful market infrastructure. But when supply is dominated by a narrow class of GPU hosts, decentralization migrates upward into payments, coordination, and governance, while the compute substrate remains concentrated. Tokenization can decentralize access without decentralizing power, and that distinction is material for users, investors, and regulators assessing operational resilience before capital allocation or integration decisions.
The industry should stop treating “decentralized AI” as a default status and start treating it as an auditable claim. Networks should publish provider concentration, maximum supplier share, hardware distribution, uptime variance, geographic exposure, pricing dispersion, and dependency on Nvidia, cloud partners, or a small validator set. That evidence should become table stakes before investors, developers, and enterprises finance the next infrastructure cycle responsibly. A credible test is simple: can the system keep serving meaningful workloads if its top providers disappear? If not, the architecture is not decentralized in the operational sense. The next AI blockchains may become useful compute markets, but many are closer to centralized supercomputers with token rails than anyone wants to admit.
