Opinion by: Leo Fan, founder of Cysic

Decentralized compute has failed. Not because it can’t find you a cheap GPU; it’s actually quite good at that. The problem is that every major network today still forces you to trust the node operator with your data and results. 

We have replaced Amazon’s login page with a wallet connection and called it Web3.

A staggering $2 billion to $3 billion was poured into “decentralized cloud” tokens from 2023 to 2025. Yet none of the top players can give a smart contract mathematical certainty that the work was done correctly. Zero-knowledge rollups, onchain AI agents and fully trustless apps remain impossible at scale.

The entire sector has decentralized supply and payments. Trust is still centralized. Until verification is cryptographic, “decentralized compute” is just Airbnb for GPUs.

The marketplace mirage

Current leaders are sophisticated spot markets, nothing more. Akash pulled in about $11 million in Q3 2025 revenue. Render managed about $18 million. Impressive for coordination layers, sure, but trivial next to AWS’s $100 billion-plus annual run rate.

These networks solved the easy part, idle GPU discovery and crypto payments, and declared victory. Their proof-of-work done? Usually, just “the node streamed the result plus some reputation score.”

That’s not verification. That’s a pinky promise with extra steps.

Real-world failures are already happening. In 2025, bad actors returned corrupted Blender renders through Render’s network. No onchain way to detect it. Io.net caught a Sybil cluster gaming reputation scores in May and further failures in November with aPriori’s mysterious Sybil cluster that claimed 60% of the airdrop across 14,000 wallets. Gensyn’s own whitepaper admits their “learning game” tolerates less than 49% malicious tolerance in practice.

These are the predictable outcomes when you replace mathematical proofs with social enforcement.

Think about what this means for actual use cases. A Layer 2 rollup outsourcing STARK proofs to any current decloud still needs a trusted multisig or single honest prover. The centralization risk remains unchanged. An autonomous agent doing inference on io.net? The on-chain contract can’t tell if the LLM output was correct or backdoored. We’ve recreated the oracle problem with more steps.

Breaking Web3’s core promise

Bitcoin never asked you to trust miners. Ethereum doesn’t require faith in validators. They gave you ways to verify. Today’s compute networks do the opposite:

“Here’s your result. Trust me, bro, and we’ll slash if someone complains.”

This philosophical mismatch kills the entire value proposition. The Total Addressable Market (TAM) for “decentralized GPU” gets capped at rendering and basic training because nobody will run sensitive workloads on networks where nodes see your plaintext data, such as DeFi bots, medical inference, and proprietary models.

Vitalik nailed it at Devcon 2024:

“If your scaling solution reintroduces trusted parties, you haven’t scaled. You’ve just outsourced.” 

That’s exactly what we’ve done. We outsourced AWS to a thousand smaller AWS nodes and patted ourselves on the back.

The market size illusion becomes clear when you do the math. Without verifiable execution, you can’t serve. Financial institutions need provable compliance. Healthcare systems require an auditable inference. Rollups demand trustless proof generation. AI agents must execute high-value transactions.

Related: Institutions must stake Ether on decentralized infrastructure

You’re left competing for Stable Diffusion hobbyists and Blender farms. Good luck building a trillion-dollar market on that.

The only path forward

Real decentralized compute requires cryptographic proof accompanying every result, including zkSNARKs, STARKs or optimistic fraud proofs, that are verifiable in under a second by any smart contract.

This isn’t theoretical anymore. Hardware-accelerated proving stacks using FPGAs and custom ASICs make this economically viable at GPU-scale bandwidth. The 2024-2025 ZPrize winners showed STARKs over cycle-accurate circuits running in under eight seconds on the latest FPGA clusters, heading toward sub-second on next-gen silicon.

When this verification layer exists, everything changes. A $10,000 DeFi agent can run private AlphaTensor-level reasoning onchain. Rollups can outsource proofs to 10,000 untrusted nodes with zero risk. Inference becomes as trustless as checking an Ethereum balance.

Open, permissionless networks of specialized provers will compete on latency and cost. But the key difference is that dishonesty becomes mathematically impossible, not just expensive. No reputation systems. No slashing games. Just math.

The real revolution

We didn’t decentralize compute by turning GPUs into an open market. That’s like saying we decentralized money by letting people trade dollars on DEXs.

We’ll deserve the name when computational results become as unforgeable as Bitcoin transactions are unspendable without the private key. It’s impossible to fake, trivial to check.

The breakthrough Web3 needs isn’t another 5% cheaper GPU hour. It’s the first network that can attach an unbreakable proof of correctness to every teraflop. That’s the infrastructure we were promised. Everything else is just a centralized cloud with extra steps.

Opinion by: Leo Fan, founder of Cysic