SIGNAL VS NOISE

DECENTRALISED
AI

OpenAI, Anthropic, and Google are accumulating unprecedented concentrations of AI capability. The response from crypto: token-incentivised networks, decentralised compute, and open model marketplaces. Here is what is real, what is narrative, and what is actually winning.

Sources: Messari, CoinDesk, Grayscale Research, io.net, Akash Network — data as of Mar 2026
"DECENTRALISED AI" IS THREE DIFFERENT THINGS
DECENTRALISED COMPUTE
io.net · Akash · Render

GPU marketplaces. Anyone's idle hardware can serve AI workloads. Genuinely cheaper than AWS. Token = payment rail and coordination mechanism. The most legitimate category.

🧠
DECENTRALISED MODEL INCENTIVES
Bittensor (TAO)

Token rewards for producing best AI outputs. Validators evaluate quality. Subnets specialise by task. The most innovative architecture. Also the furthest from frontier capability.

OPEN WEIGHTS / OPEN SOURCE
Llama · DeepSeek · Qwen

No token. No network. Just model weights that anyone can download and run. Arguably the most powerful decentralisation of AI capability ever achieved. And it's free.

NETWORK DEEP DIVES

io.net (IO)

Decentralised Compute · Founded 2023
GENUINELY DECENTRALISED

Decentralised GPU cloud. Aggregates idle GPUs from data centres, mining farms, and consumers into on-demand clusters.

MARKET CAP
~$500M–1B
ACTIVE NODES
320,000+ verified GPUs
COMPUTE PRICE
$0.76–1.93/hr for A100–H200 (vs AWS $3–8/hr)
VS AWS
70–80% cheaper (benchmark; actual varies by workload)
HOW IT WORKS

Deploy a 10,000-GPU cluster in under 10 seconds. Uses Solana for payments. Supports Kubernetes workloads. 70% cheaper than cloud providers in benchmarks.

Source: io.net blog 2025, Messari, blocmates Aug 2025
REAL USAGE

Real: Wondera scaled AI music creation to 200K users using io.net, cut training costs 75%. Multiple AI startups using it as AWS alternative.

CAPABILITY GAP / LIMITS

Not about model capability — provides raw compute. Same GPU = same performance regardless of who owns it.

KEY RISKS

GPU hardware verification is hard — fake/low-spec nodes are a known attack vector. Sustained utilisation rates not disclosed.

THE UNCOMFORTABLE TRUTH

Meta Did More to Decentralise AI With One GitHub Repo Than All Crypto Networks Combined

Llama 4 Scout (109B) and Maverick (400B) are free. DeepSeek R1 671B is MIT licensed. Qwen 3 235B runs on a Mac Studio. These models match or approach frontier capability and anyone on Earth can download and run them. That is decentralisation of AI capability. Token-based networks are about economic decentralisation — who captures value. They solve a different problem. Both matter, but they are not the same thing.

KEY OPEN WEIGHTS MODELS — WHAT YOU CAN RUN WITHOUT PERMISSION
MODEL
ORG
PARAMETERS
LICENSE
WHAT IT DECENTRALISES
Llama 4 Scout
Meta
109B (MoE)
Open weights
Capability — anyone can run frontier-class model locally
Llama 4 Maverick
Meta
400B (MoE)
Open weights
Full model for fine-tuning and deployment — no API dependency
DeepSeek R1
DeepSeek
671B (MoE)
MIT
Reasoning model matching o1 — free, no rate limits, self-hostable
Qwen 3 235B
Alibaba
235B (MoE)
Apache 2.0
Multilingual frontier model available to anyone globally
Mistral Large 2
Mistral
123B
Research
European frontier model, API + self-hostable, no Big Tech dependency
Gemma 3 27B
Google
27B
Apache 2.0
Consumer-runnable capable model — fits in 32GB RAM on a Mac
DECENTRALISED COMPUTE vs HYPERSCALERS — GPU PRICE COMPARISON
GPU
AWS
AZURE
IO.NET
AKASH
SAVINGS VS AWS
NVIDIA A100 80GB
$3.40/hr
$3.67/hr
~$0.85/hr
$0.76/hr
78–80%
NVIDIA H100 80GB
$8.00/hr
$8.50/hr
~$2.10/hr
~$2.50/hr
69–74%
NVIDIA H200 141GB
~$12/hr
~$13/hr
Limited avail.
$1.93/hr*
84%* (Akash)
NVIDIA RTX 4090
N/A (consumer)
N/A
~$0.45/hr
~$0.50/hr
vs equivalent: 60–70%
*H200 on Akash reflects Q1 2025 pricing. Availability limited. AWS/Azure figures approximate on-demand rates. Sources: Messari Q1 2025, Coin Bureau, io.net.
THE HONEST VERDICTS
Decentralised Compute (io.net, Akash, Render)
LEGITIMATELY USEFUL

Real price advantage (70–83% cheaper than AWS), real usage, real customers. The token is a payment/coordination mechanism. The underlying service is genuine.

Decentralised Model Incentives (Bittensor)
INTERESTING EXPERIMENT

The incentive architecture for collective intelligence is genuinely novel. 128 subnets with Darwinian selection is a real innovation. But the capability gap vs frontier labs remains significant. Watch, don't write off.

Open Source AI (Llama, Qwen, DeepSeek)
THE REAL DECENTRALISATION

Meta released Llama 4 and did more to decentralise AI capability than all crypto tokens combined. Anyone on Earth can run a frontier-class model. No token. No fees. That's the actual distribution of power.

AI Governance Tokens (most others)
MOSTLY NARRATIVE

Adding a token to a centralised AI product does not decentralise it. "Owning" a governance token is not the same as owning model weights. Many projects conflate the two. Buyer beware.

THE POWER BALANCE QUESTION

Who Controls AI? And Can Anything Actually Change That?

THE CENTRALISATION CONCERN IS LEGITIMATE

77% of consumers in a 2025 Harris Poll said decentralised AI is more beneficial than Big Tech-controlled systems. OpenAI, Anthropic, and Google have unprecedented concentrations of capability, capital, and data. The Dario/Pentagon case showed a single CEO can set AI red lines for the US military. That is real power concentration.

CRYPTO'S ANSWER IS PARTIAL

Decentralised compute (Akash, io.net) genuinely lowers the cost barrier and removes single-provider dependency. Bittensor's incentive architecture is a real experiment in collective intelligence. But none of these systems currently produce AI at the capability level of GPT-4o or Claude Sonnet. The gap is the thing to watch.

OPEN SOURCE IS WINNING THE CAPABILITY RACE

Llama 4 Maverick runs on a Mac Studio. DeepSeek R1 is MIT licensed and matches o1 on reasoning. The open source movement has done more to democratise frontier-level AI capability than any token network. The question is whether open source can stay within 6–12 months of frontier indefinitely.