Akash, io.net, Render — and the gaps we exploit · 07_compute_marketplace/COMPUTE_MARKET_COMPETITORS_2026.md
Compute Market Competitors — 2026 State of the World
Date: 2026-09-08 · Author: qalcode autonomous research+build session
Method: DuckDuckGo/Bing web search snippets fetched 2026-09-08 (see dated citations inline). Prices/metrics are from the cited secondary sources, not primary dashboards — anything I could not corroborate from at least one dated 2026 source is tagged [UNVERIFIED]. This doc feeds P2P_COMPUTE_MARKET.md and the qmarket v2 build (stakes/slashing/reputation, remote providers, bench oracle).
1. Why this scan exists
Qmarket v1 (2026-09-06) proved two-party compute sales in QALS-backed credit with on-chain provenance. Before building v2 (provider-held stake, dispute→slash, reputation, remote provider agents), we checked what the nine serious players do for payment unit, escrow/trust, provenance, and cold-start — and what they still get wrong in September 2026.
The one-line headline: nobody in the sector pays providers in a stable, redeemable unit, and nobody anchors output provenance on-chain. Trust is stuck at "trust the operator" — none of the nine has per-provider stake → dispute → slash → delist tied to output correctness. That is exactly the v2 gap.
2. The field
Akash Network (Cosmos → post-Cosmos, 2026)
- Model: Reverse-auction marketplace; providers bid, tenants pay the winning per-GPU-hour bid (USD or AKT). H100/H200/A100/RTX-class across 150+ providers, 30+ countries; consumer cards from ~$0.03–0.05/hr (gpus.io, clusterbid.com, 2026). Third-party analyses claim 60–85% below hyperscaler pricing (cryptoaianalysis.com, 2026-06-30).
- Payment unit & volatility: AKT ≈ $0.73, mcap ≈ $215M (kkdemian.com, 2026-06-23). Down from $6+ 2021 highs → sellers price in USD and hedge, i.e. the token is a settlement detail, not the pitch. 2026 stories describe a "Cosmos exit" and ACT settlement + BME burn changes (hoge.gg, 2026-08-24) [UNVERIFIED details].
- Escrow/trust: Deployment escrow via Akash's chain (bid→lease→close lifecycle is the cleanest in the sector). Provider telemetry is largely trusted; no output correctness verification, no slashing for wrong results.
- Provenance: none. Deployment manifests are on-chain; results are not.
- Cold start: Utilization lags the AI hype (hoge.gg, 2026-07-15); Messari Q1-2026 revenue read is the honest demand measure [UNVERIFIED figure].
- Copy: the bid/lease/close lifecycle and provider self-describe SDL manifests (→ future qmarket
POST /providersbid fields).
io.net (Solana DePIN)
- Model: Aggregate consumer+DC GPUs into clusters for ML training/inference. In June 2026 launched the Incentive Dynamic Engine (IDE): IO emissions/burns tied to real network earnings, ≥50% of post-payout revenue burned (dextools.io 2026-06; CoinMarketCap 2026-09-07). Daily network earnings ≈ $35–36K/day trending up since March 2026 (cryptobreaking.com citing CoinDesk Research, 2026-06-17) — across an entire "largest decentralized GPU network", that is remarkably small. $8M enterprise deals + 12M IO burn (ainvest.com, 2026-06-14).
- Payment unit & volatility: IO ≈ $0.109 (gpumarketdepin.com, 2026-02-21); ~$3–7/day per RTX 4090 claimed [UNVERIFIED]. IO is down >95% from 2024 highs — the volatility problem in one number.
- Escrow/trust: central operator (io.net the company) vets workers; "Proof of Compute" telemetry from their agent — trusted client-side. No stake/slash per output.
- Provenance: none on results.
- Cold start: subsidy-heavy emissions; Agent Cloud (March 2026) lets AI agents autonomously rent GPUs and pay in stablecoins (bex.co, 2026-05-03) — validation that agent-buyers are real, and that they want stable payment units, which io.net answers only at the edge.
- Copy: cluster abstraction (job→cluster→worker routing) and the agent-autonomous-buyer pattern (qmarket buyers are literally apps/agents).
Golem Network (Ethereum, GLM)
- Model: The OG. Requestor/provider daemon market for idle CPU (Blender, rendering, science), slow pivot to AI/GPU. GLM ≈ $0.11, mcap ≈ $110M (OKX/CoinDesk, 2026-09-08).
- Payment unit & volatility: GLM — an ERC-20 that has traded ~$0.09–0.35 for five years. Low velocity of actual usage: "public project data does not yet provide clear evidence that commercial usage and paid workloads rose at the same pace as the token price" (tapbit.com, 2026-08-31).
- Escrow/trust: on-chain deposit + task agreement; CPU tasks get concentration-based consensus (N providers compute the same task, majority hash wins) — the sector's only real recompute-verification, but it is (a) CPU-task-specific, (b) expensive (N× compute), and (c) absent for GPU/AI outputs.
- Provenance: none.
- Cold start: 10 years old and still niche — the cautionary tale for token-first marketplaces without an anchor tenant.
- Copy: deterministic-recompute dispute checks (our v2
/disputerecompute path is Golem's consensus idea, sampled and cheap).
Render Network (Solana, RENDER)
- Model: Two-sided GPU render + AI job market (Octane heritage). Burn-Mint Equilibrium: jobs priced in USD, buyer's RENDER burned at payment, node operators paid from emissions. Actively pivoting from 3D render to AI compute (web3wagmi 2026-07-05; hoge.gg 2026-06-26).
- Payment unit & volatility: RENDER ≈ $1.68, mcap ≈ $872M (kkdemian.com, 2026-06-22). Users are insulated (USD-priced jobs) but operators are not: payouts arrive in a token that fell from $13.60 ATH to ~$1.68.
- Escrow/trust: job escrow via network payments; node trust is reputation + spot-checks by the Foundation [UNVERIFIED mechanism detail]. No on-chain output proofs.
- Provenance: burn records are on-chain; output hashes are not anchored — the burn proves payment, not result.
- Cold start: the key datapoint — July 2026 flow analysis: ~1.48M RENDER minted vs ~0.18M burned over 90 days (≈8:1 mint:burn) (mrnasdog.com/research/render/inflation, 2026-07). Subsidies are 8× real demand. Cold start is being bought, not solved.
- Copy: burn-mint equilibrium as a treasury sink (our gateway fee → FloorVault is the paper-phase ancestor; on-chain phase can burn a slice).
Nosana (Solana, NOS)
- Model: CI/CD-pivoted AI inference grid; consumer RTX boxes run containerized inference jobs ("Nosana Grid"); lowest-latency niche (depin.builders report, 2026-05-29). NOS ≈ $0.20 (decentralizedinference.org, 2026-02-04).
- Payment unit & volatility: NOS — Solana token, governance + settlement coordination (nosana.com).
- Escrow/trust: job fees escrowed per job run; node vetting light; relies on container determinism and market reputation [UNVERIFIED].
- Provenance: none.
- Cold start: rides the same "cheap inference" story as everyone; token ran up on speculation (genztech.blog, 2026-07-13).
- Copy: containerized-job simplicity — one spec, one container, one price (our provider.py /execute is the same shape).
Fluence (FLT)
- Model: Enterprise DePIN from Tier 3/4 data-centers (not consumer boxes); GPU compute GA since Oct 2025 (fluence.network, 2025-10-03), public access March 2026 (Fortune press release, 2026-03-12). On-chain revenue records; DAO-governed.
- Payment unit & volatility: FLT for staking/governance/coordination — and notably an FLT-collateralized stablecoin for settlement (depinspace.co, 2025-10-20). That is the sector's second stable-unit experiment after Render's USD-pricing — evidence the market is converging on our thesis.
- Escrow/trust: provider SLAs, DAO slashing for infra providers [UNVERIFIED detail]; enterprise contracts do the trusting.
- Provenance: on-chain revenue, not outputs.
- Cold start: solved the boring way — data-center partners with real enterprise deals.
- Copy: on-chain revenue accounting (loopd receipts are the paper version) and the stablecoin-settlement instinct.
Aethir (ATH)
- Model: Enterprise GPU cloud (AI + gaming) with distributed hosts; Checker Node layer — $99-license nodes that verify host GPU specs and QoS (app.aethir.com; depinrun.com guide, 2026-04-28). >$166M annualized revenue claimed (CoinMarketCap news, 2026-08-21) — the sector's biggest real revenue [UNVERIFIED].
- Payment unit & volatility: ATH; Coinbase delisted ATH perps 2026-09-03 (CMC news, 2026-09-07). Supply reporting "diverges", cash flow doesn't clearly reach token holders (kkdemian.com, 2026-06-28).
- Escrow/trust: Checker is the interesting bit — decentralized verification of hardware, but of specs, not outputs. Hosts stake ATH; penalties for downtime.
- Provenance: none.
- Cold start: bought with a huge token float and enterprise BD.
- Copy: the Checker pattern → our
bench.pyis the honest version: measure actual performance, don't just attest specs.
Bittensor compute subnets (TAO)
- Model: Not a marketplace — an emissions game. Subnets (80–129+ active in 2026 reports; capacity doubled toward 256) run validators who score miners by re-running inference; dTAO gives each subnet an alpha token; SN64 "Chutes" is the dominant crypto-AI inference marketplace (defi-intel.com 2026; coingecko.com 2026-05-07). TAO ≈ $229, subnet mcap ≈ $1.12B (taonews.ai, 2026-09-08).
- Payment unit & volatility: TAO + subnet alphas — the most volatile units in the field; miners are effectively TAO-weather farmers.
- Escrow/trust: the validator re-scoring loop is the escrow — closest cousin to our dispute/recompute, but incentives are emissions-share, not per-job payment; gaming (fake quality, weight-copying) is a chronic subnet problem; "Covenant AI exit" exposed governance risk (gate.com, 2026-05-25).
- Provenance: none.
- Cold start: solved by emissions (subsidy), same as Render.
- Copy: validator-style re-inference scoring (bench.py + dispute recompute), while avoiding emissions-subsidy economics entirely.
Morpheus / agent-compute plays (MOR)
- Model: "A network for powering smart agents" — MOR20 fair-launch, agents that never hold private keys, and the Lumerin Node compute marketplace (router contracts over peer GPU endpoints) (github.com/MorpheusAIs/Docs, fetched 2026-09-08).
- Payment unit & volatility: MOR (ETH/Arbitrum/Base). Docs repo shows ~80 stars — momentum has visibly faded vs the 2024 peak [UNVERIFIED sentiment read].
- Escrow/trust: smart-contract router holds payment until compute contract completes [UNVERIFIED detail]; agent-safety principles (no keys) are genuinely good.
- Provenance: none.
- Cold start: community-subsidized; largely stalled.
- Copy: agents-as-principals (our buyers are apps/agents with loopd agents + HMAC keys — same shape, shipping today).
3. Gaps we exploit
| # | Gap in the field (2026-09) | Qmarket v2 answer |
|---|---|---|
| 1 | Every network pays in a volatile token (AKT $0.73, IO $0.11, GLM $0.11, RENDER $1.68, NOS $0.20, TAO $229) — sellers wear FX risk on top of compute risk | B-QALS credit, AU$1:1 redeemable via loopd holds — seller price risk is zero |
| 2 | No output provenance anywhere — Render burns prove payment, not results; nobody anchors result hashes | Every job's output sha256 anchored on Qalnet via qalpipe; verify is a chain read |
| 3 | No stake→dispute→slash→delist per provider — Aethir Checker checks hardware, not honesty; Golem consensus is CPU-only; Bittensor scores emissions not jobs | v2: stake_cents lock, /dispute with deterministic recompute + integrity check, 50/50 slash to treasury(FloorVault)/buyer, reputation −25, delist <40 |
| 4 | Reputation is marketing, not a number | reputation is persisted, math-readable, buyer-filterable (/providers excludes delisted) |
| 5 | Cold start is bought, not built (Render mint:burn ≈ 8:1; io.net network-wide earnings ≈ $35K/day) | qalarc apps are anchor tenants — tulpa/tradez/hub buy first, subsidies stay at zero |
| 6 | Pricing is opaque (bid wars on Akash, emissions games on Bittensor) | bench.py price-floor oracle: measured sha256/tok-per-s → target-margin suggested cents/unit, in the open |
| 7 | Providers are servers-in-someone-else's-cluster (io.net/Golem daemons) | provider.py: the seller runs their own zero-dep agent on their own port, registers with the gateway, keeps their adapter |
4. What we COPY (concretely)
| From | Copy into qmarket | Status |
|---|---|---|
| Akash bid→lease→close lifecycle | job states + hold→settle→release mapping | ✅ v1 (loopd holds = lease escrow) |
| io.net Agent Cloud | buyer = agent with HMAC key, autonomous purchase | ✅ v1 (loopd agents) |
| io.net cluster abstraction | provider advertises adapter URL; gateway routes | ✅ v2 (adapter_url) |
| Render burn-mint | gateway fee slice → treasury FloorVault (future: burn) | 🟡 paper (10% fee) |
| Golem recompute consensus | /dispute deterministic recompute |
✅ v2 |
| Aethir Checker | bench.py measures real perf, not attested specs | ✅ v2 |
| Bittensor validator scoring | recompute as reputation input (+1/−25) | ✅ v2 |
| Fluence stablecoin settlement | B-QALS stable unit thesis confirmed | ✅ thesis (already shipped) |
| Morpheus agent principles | agents never hold buyer keys; gateway signs | ✅ v1 (HMAC pattern) |
5. Verdict
The 2026 field is big tokens, thin trust: real demand remains a rounding error next to the subsidies (Render 8:1, io.net $35K/day), units are volatile, and no one verifies that the compute you bought actually happened correctly — or can prove afterwards what was produced. Qmarket's wedges — stable redeemable credit, per-job on-chain provenance, and now stake-backed reputation with recompute disputes — are each unique in the set. v2 builds all three.