Decentralized Compute Tuesday — September 1 2026 Weekly Update

Weekly covering August 26 – September 1, 2026. Three structural threads define the week. First: the Stripe/OpenRouter era begins to show its shape — OpenRouter's public provider page now lists Phala at 24.0B tokens/week (496.9B total) and AkashML at 23.6B/week (489.1B total) as the top-ranked decentralized providers, putting both within rounding distance of Groq (20.7B/week). The new gatekeeper's directory is the de facto product surface for every DePIN inference network. Second: TAO's spot-ETF decision window remains open and Grayscale amended its S-1 (Aug 31) — the procedural move that signals SEC engagement rather than dismissal. Third: dstack went cross-cloud — Phala's Zero-Trust confidential container runtime now runs on AWS Nitro Enclaves and Google Cloud Confidential VMs with the same measurement, KMS, and gateway stack as on Phala's own network; the architectural argument for confidential compute portability moved from design doc to working code. Below: ACCELERATE tokenomics refresh, Bittensor on Base (CCIP) settling, Gensyn's 100-day retrospective (Delphi thriving, RL Swarm paused), and a Phala spike on the new "Qwen3.7 Flash" Chinese open-weight family.

Macro Thread — The Stripe/OpenRouter Provider Page Is Now the DePIN Inference Dashboard

Until late August the only third-party measurement of DePIN inference throughput was the DeAI Dashboard (covered Aug 25). With the Stripe acquisition of OpenRouter, the OpenRouter providers page itself became a live, public ranking of every network selling through it:

What it tells you:

1. The two decentralized providers ranked #5 and #6 by last-week tokens — Phala and AkashML each above Groq in fresh token volume. Across the lifetime column, Venice (centralized), Cohere, Friendli, and Groq have shipped more; Phala and AkashML are inside the top quartile of all listed providers.
2. Phala's 21 listed models outnumber AkashML's 6 — the dstack confidential catalog (GLM-5.2, DeepSeek V4 Flash, Qwen3.7 Flash, Qwen3 Max, Kimi K2.6, Llama 3.x, Phi-3, Muse Glimmer, etc.) is now the deepest confidential-AI menu on OpenRouter, and it ships under a single "zero retention" privacy contract.
3. AkashML is leveraged on volume per model — six models producing 23.6B tokens in a week means AkashML is averaging nearly 4B tokens/model/week (heavily skewed by DeepSeek V4 Flash). Concentrated demand on fewer open-weight SKUs is the explicit AkashML thesis ("the Kubernetes of inference at AWS-replacement prices").
4. The governance implication: a single Stripe-owned dashboard sets the default discoverability surface for DePIN inference demand. Network strategy teams should treat the OpenRouter provider page as their landing page. Listings, model count, weekly tokens, and "Zero retention" badges are now the marketing surface — not blog posts.

Sources: OpenRouter providers (live rankings snapshot Aug 31 2026); CoinMarketCap Akash and CoinMarketCap Phala for narrative context.

Phala — dstack Goes Cross-Cloud; Phala Network OpenRouter Catalog Hits 24B Tokens/Week

Cross-cloud dstack: AWS Nitro, GCP TDX, Phala native, with one measurement stack

Phala's dstack whitepaper (Aug 2026) and the One Confidential Compute Framework post document the same workload definition now packaging into three runtimes:

Developer: docker-compose + signed manifest
                │
                ▼
   ┌────────────────────────────┐
   │ dstack (workload identity, │
   │ KMS, gateway, attestation) │
   └────────────┬───────────────┘
                │
   ┌────────────┼──────────────┐
   ▼            ▼              ▼
Phala TDX   GCP TDX+TPM    AWS Nitro
(native)    (EIF/CCV)      (Enclave)
                │              │
                └──────┬───────┘
                       ▼
               Per-app keys
               (released only after
                measurement-verified
                attestation passes)

Three architectural moves make this non-trivial:

• Portable confidential containers: workload portability across TDX / SEV-SNP / Nitro without rewriting the compose file. The "last-mile" packaging step (EIF construction on Nitro, CCV image build on GCP) is owned by the platform adapter, not the developer.
• dstack-KMS on-chain: keys are bound to a measurement-derived app identity, not a cloud account. Forward/backward secrecy with key rotation; multi-party computation supported for distributed trust.
• Governance collusion protection: dstack's policy smart contracts require multisig (default 2/3) and a timelock delay before code swaps can release new keys — operators cannot silently swap binaries or exfiltrate secrets.

The threat model is the cloud operator, not the silicon vendor. dstack's position is that the silicon (Intel TDX, AMD SEV-SNP, NVIDIA Hopper/Blackwell CC, AWS Nitro) is the necessary root of trust, but policy enforcement and identity binding must live above it. This matches the OYM (Aug) reading of the TEE landscape.

OpenRouter trajectory and the 21-model catalog

Phala now serves 21 models on OpenRouter, zero-retention, US-jurisdiction, with a 24B-token week ending Aug 31. The active model mix (per official @PhalaNetwork breakdowns):

• DeepSeek V4 Flash — still the single biggest SKU (~30%+ of weekly volume since Aug 22).
• GLM-5.2 — Zhipu AI's flagship, surging on the Aug 14–28 China-US trade-policy backdrop (Anthropic's export-control compliance move pushed Chinese-origin models into the OpenRouter enterprise channel per CNBC Jul 7; Phala's confidential hosting makes them audit-friendly).
• Qwen3.7 Flash / Qwen3 Max — the Alibaba open-weights family added to the dstack catalog; Qwen3.7-Flash hit ~1M OpenRouter requests the week it shipped.
• Kimi K2.6 / GLM-5.x 30B / Muse Glimmer — the open-weight refusal-friendly cluster that lets Phala bypass US political content-moderation defaults.
• Llama 3.x / Phi-3 14B-128k / DeepSeek V3 / Mistral / Gemma — the standard set.

The honest counter-signal: PHA at ~$0.022, 99% off its 2024 high, while billed tokens have grown ~7x in three weeks. Token value-capture from confidential-inference volume remains the unsolved design problem in the sector.

Performance proof: TEE-on vs TEE-off deltas are tiny on H100/H200

The Phala performance report now publishes side-by-side TEE-on / TEE-off benchmarks:

The 70B numbers — zero or near-zero overhead — are the strongest confidential-inference proof point in the sector. The dstack architecture's per-app-key + KMS + gateway path is cheap when the model's already in HBM.

Sources: Phala dstack page, dstack whitepaper, performance report, @PhalaNetwork, OpenRouter providers.

Bittensor — Grayscale S-1 Amendment Filed (Aug 31); Chutes Plateau; Exploit Summit 4 Weeks Out

Grayscale amended its S-1 on August 31

The procedural signal most market participants were watching: Grayscale filed an amended S-1 for the Grayscale Bittensor Trust ETF (NYSE Arca, ticker GTAO) on Aug 31, per Cryptorank coverage. An amendment during the review window is interpreted two ways:

• Bullish reading: SEC engagement continues; amendments typically signal the regulator is asking specific questions and Grayscale is responding. Aug 31 is also squarely inside the originally-forecasted "August 2026" decision window.
• Bearish reading: amendments are routine; the underlying regulatory obstacles (no CME futures on TAO; manipulation-resistance concerns; surveillance-sharing agreements) remain unresolved. Per Binance Square analysis, explicit approval odds estimated at ~35% before end-2026 / ~50% sometime in 2027 / ~15% further delays.

Grayscale's TAO allocation inside its AI fund held at 43% (largest position) in the August update, up from 29.15% at Q2 rebalance — the manager keeps adding exposure into the binary outcome. TAO held ~$200–232 through the week; mcap estimates $2.2–2.6B.

TAO on Base via CCIP — flow ramps

The Base/CCIP/Aerodrome bridge covered Aug 25 continues to attract liquidity. The TAO-USDC pool on Aerodrome is now eligible for AERO emissions, and Bittensor's subnet-asset listings on Kraken (Jun) plus the Base distribution layer are the deliberate cross-L2 strategy. The structural read: the Bittensor community is positioning TAO for the same retail-distribution rails that made SOL and ETH post-2020 ETF stories tractable.

Chutes plateau and Chutes-specific dynamics

Per the DeAI Dashboard (Aug 14 snapshot, the most recent third-party count): Chutes SN64 logged 425.35B tokens/day — by far the largest single network. Through August the trajectory has been roughly flat (no new subnets have yet replicated the Chutes surge). The dashboard's silence on subnet churn since Aug 14 means either Chutes has stabilized or measurement hasn't refreshed; the dashboard update cadence matters.

Teutonic-II (Subnet 3, Const) — 110B MoE training run

The largest permissionless training event in Bittensor's history launched the previous fortnight. Teutonic-II is a 110B-parameter sparse MoE that scales the competition-based Teutonic mechanism from 10B — direct test of "can Bittensor train frontier-scale models in a subnet?" If it converges, it becomes the subnet template that other pretraining-focused subnets (Subnet 9 historically) would copy.

Exploit Summit (Sept 28–29, Montreal, New City Gas)

The flagship Bittensor gathering is now four weeks out:

• Date: Sept 28–29 2026.
• Venue: New City Gas, Griffintown, Montreal (restored 19th-century industrial complex).
• Organizer: Bittensor Commons (non-profit).
• Confirmed speakers: Jacob Steeves (Affine CEO, Bittensor co-founder) + ecosystem founders across 100+ subnet teams.
• Format: 2 days of debates, live subnet demos, workshops. Adversarial by design — "Game on for game-theoretic AI" is the tagline.
• Media partners: AMB Crypto, CoinTelegraph, TAO Times, TAO Daily.
• Tickets: Pre-sale, Early-bird, Main release (sold out by Aug 10), and Student tickets still open (university email + ID required).

The substantive agenda items shaping up: dTAO underwriting under ETF context (Mark Creaser's "rule changes make dTAO harder to underwrite" thread), post-Covenant validator-competition restoration progress, subnet consolidation debates, and the inaugural subnet-teams showcase.

Sources: Grayscale amended S-1 / Cryptorank Aug 31, Binance Square ETF analysis, Exploit Summit 2026, Luma ticket page, TAO Daily preview.

Akash — AEP-46/67 Deliveries; 23.6B OpenRouter Tokens/Week; Homenode Update

AEP-46 (Spot/Preemptible Instances) and AEP-67 (Console Bid Screening) land

Both roadmap items due Aug 30 (AEP-46) and Aug 31 (AEP-67) cleared the deadline window:

• AEP-46 — Spot Instances / Reserved Instances: providers can now whitelist tenant wallets (reserved) and offer discounted interruptible capacity (preemptible). Two-sided enterprise feature set — addresses the largest objections from AWS/Google Cloud workloads about bid predictability and lease longevity.
• AEP-67 — Console Bid Screening: Console pre-screens providers against a deployment's requirements, so users see only providers who would actually bid. Replaces the abandoned live per-provider inventory query design. Zero-bid-UX is a direct response to the Console friction cited in Aug 4.

The AEP-85 (Buildpacks-Powered Deployments, Sept 30) is next — replaces the broken Build & Deploy flow (hard-coded JS frameworks, broken for Python/Go/Rust/.NET) with Cloud Native Buildpacks. That lands before the Oct 3–5 Akash Annual Conference + Hackathon in SF — a likely venue for L1/shared-security RFP (AEP-79, targets Dec 31) news.

OpenRouter position: 23.6B tokens/week, 489.1B lifetime

AkashML is now the #2 decentralized provider on OpenRouter (behind Phala) at 23.6B tokens/week — making AkashML and Phala the only DePIN networks in the OpenRouter top 10 by recent throughput. AkashML's narrower catalog (6 models) but heavier per-model volume shows the "Kubernetes-of-inference at AWS-replacement prices" thesis is now landing measurable demand, not just capacity.

Homenode and the GPU supply gap

The Homenode Beta (consumer/prosumer GPUs) covered in previous weekly remains in testing. The supply-side gap — Messari's 58 providers avg, network low in Q1 — is the binding constraint on Akash growth, and Homenode is the team's experimental answer. Akash Insiders are stress-testing provider onboarding, hardware compatibility, and inference performance ahead of the alpha release. Watch the October conference for beta-to-alpha graduation.

Sources: AEP-46 Spot Instances, AEP-67 Console Bid Screening, Akash 2026 roadmap, OpenRouter providers, Akash AI apps blog.

Aethir — ACCELERATE Tokenomics Refresh + Axe Compute Q2 Close-Out

Two cleanups from the Aug 24 ACCELERATE announcement (covered Aug 25):

IDC tokenomics refresh is incoming

The Aug 24 announcement bundled ACCELERATE with a tokenomics policy update: a burn mechanism plus a variable platform fee. The intent — per the original Aethir post — is to make ATH capture on-chain revenue during price volatility. This is the cleanest, most explicitly-stated "burn-on-real-revenue" tokenomics in DePIN compute; watch for the actual on-chain emission schedule when it ships.

Axe Compute Q2 — $317M prepayments at the line

Axe Compute's Q2 close-out (Aug 18, SEC-reporting neocloud CIK 0001446159):

• $317M+ customer prepayments received on top of $60.8M at June 30, against the >$3B July contracts.
• $3.2B signed 2026 contract value; $696M run-rate at full deployment vs $37M exiting Q2.
• $5.9B qualified pipeline across 98 opportunities.
• Duos Technologies deal: ~49% project equity in 55 MW of US capacity for $140M.
• Columbus GA: empty building → 288-node Blackwell cluster in ~45 days, $20–21M quarterly revenue fully live.

The financing-model structural innovation: customers prepay 20–45% of project cost, the contracted revenue stream is used to raise project financing — infrastructure funded by its customers rather than the public market. This is the cleanest "DePIN-branded infrastructure can clear nine-figure contracts" demonstration in the sector.

Counter-signal: ATH perp delist Sep 3

Coinbase's Sept 3 ATH perpetual futures delisting (one of nine routine-contract terminations) is a non-event for spot, but the timing — contracts at $3.2B signed, ATH at ATL — sharpens the sector's defining token-vs-equity divergence.

Sources: Aethir ACCELERATE, Axe Compute Q2, Duos/GlobeNewswire.

io.net — 71.45B Tokens/Day; Wondera 200K-User Case Study Now the Sales Anchor

io.net's DeAI Dashboard position held: 71.45B tokens/day, B200 index 402.20 (highest standalone), placing it #3 after Chutes (425B) and Gonka (72B). The zkTFLOPs-verified aggregation model continues to produce measured output rather than listed capacity.

The Wondera validation is now the canonical io.net case study: AI-music startup scaled to 200,000 users across 171 countries in 4 months, consumed 552,000 GPU hours across 8-node clusters (96 H100/H200 peak), saving $2.48M vs AWS ($1.24M vs $3.72M equivalent). Wondera is now building a B2B agent for music labels (derivatives from production music — "pure AI has no copyright") and an "agent music player" — voice-activated, no playlists. The case study's revenue impact is what makes it sticky for io.net's enterprise pitch: real customers, real ARR, real cost arbitrage.

The Aug 11 unlock absorbed cleanly: 13.29M IO (~$1.66M, 1.7% of supply) released across community/seed/Series-A/core-contributor tranches with no visible dislocation; the IDE-burn mechanic continues compounding.

Editorial posture sharpened across August: "The GPU crisis won't be solved by building more data centers" (Aug 13), data sovereignty in training geography (Aug 7), and "Who Decides What Your AI Can Say?" — io.net positioning the censorship-resistance argument explicitly, alongside cost.

Sources: io.net blog, Wondera case study, CoinGecko, CMC io.net.

Gonka — Sprint Stabilization Continues; DeepSeek V4 Flash Live; Bitfury Backing Holds

Per the Gonka network-updates feed:

• Aug 14: DeepSeek V4 Flash activated (chain epoch 360), joining MiniMax-M2.7 and Kimi K2.6 as active model groups. PoC weight earned in the DeepSeek group now converts into consensus weight at the calibrated coefficient 0.214.
• Aug 20–21: Proposal 96 ("Stabilize in-epoch invalid SPRT on small samples") passed — adjusts how the invalid-inference check behaves on small samples so it doesn't misfire early in an epoch before there's enough data. Engineering hygiene vote, no host action required.

The Sprint rewrite (post Jan 2026 host-capacity-exaggeration exploit) remains in flight. The Aug governance rhythm — small, surgical, well-documented proposals — is consistent with a team shipping under audit pressure, not one ignoring the rewrite. Track this for 4–6 weeks per Aug 25's recommendation.

DeAI Dashboard position held at 72.34B tokens/day on 626 GPUs (≈$0.000314/M tokens), #2 tracked entity overall, #1 standalone. The first credentialed useful-PoW-for-AI thesis at measurable scale.

Sources: Gonka network-updates, Gonka whitepaper, crypto.news profile.

Render Network — Dispersed Power Hour Held; Throughput Question Persists

The Dispersed Power Hour (Denver, Aug 26, 22:00 UTC) — the one-hour micro-hackathon to deploy on the Dispersed AI subnet with Ollama/ComfyUI — ran as scheduled. Builder-engagement signal is positive (consistent with the post-RenderCon push: live MCP integrations for Blender/OctaneRender/Dispersed; RNP-023 Salad subnet with ~60K GPUs).

The dashboard gap still stands. No measurable inference throughput recorded for Render/Dispersed yet, against a ~$739M market cap — the highest value-per-verified-token gap in the tracked field. The creative-rendering business (seven years, film-grade) remains the real revenue engine. Watch whether the September dashboard refresh shows Dispersed throughput — that is the single most consequential number for the Render thesis.

Sources: Render Network, Dispersed, Render Foundation, CMC Render.

Prime Intellect — Autonomous-Research Benchmark; INTELLECT-3 + SYNTHETIC-2 Confirmed

Prime Intellect maintained its release-per-week cadence through August, with research output landing on the homepage:

• Aug 7 — Multi-Agent Systems in PRIME-RL (primeintellect.ai): a multi-agent extension of the RL stack, the training substrate for INTELLECT-3.
• Aug 13 — Prime Flash MoE (primeintellect.ai): Blackwell-optimized MoE kernels. Concrete GPU-kernel work, not a roadmap slide.
• Aug 14 — Measuring Autonomous AI Research (primeintellect.ai): a benchmark direction for autonomous research agents — the kind of evaluation primitive that, if adopted, becomes sector infrastructure.

INTELLECT-3 (106B total / 12B active MoE, AIME 2024/2025 at 90%) and SYNTHETIC-2 (4M collaboratively generated reasoning traces, 1,250 GPUs over 3 days from 4090s to H200s via pipeline-parallel distributed inference) are now baseline context for the project. The Sept expectation: more autonomous-research benchmark refinements, more multi-agent harness work, and Zapier/Ramp customer-story expansions.

The homepage's idle-capacity math — 6.7M idle GPU-hours resold, $20M+ potential profit at $8/GPU-hr — remains the cleanest articulation of the DePIN-training thesis anywhere in the sector. NVIDIA collaboration + Intel Capital July investment make Prime Intellect the unusual decentralized-training lab with both chip giants' fingerprints.

Source: primeintellect.ai; INTELLECT-3 + SYNTHETIC-2 retrospective.

Gensyn — 100-Day Retrospective: Delphi Thrives, RL Swarm Paused

A detailed 100-day retrospective on Gensyn (Aug 3, still the freshest substantive analysis) crystallizes the post-mainnet state:

Delphi — the live application

The AI-settled information market is Gensyn's first mainnet app. It is generating real fees and is the only layer of Gensyn's stack currently producing measurable on-chain activity. Delphi's Buy-and-Burn mechanism activates on May 1, 2026 — protocol fees automatically purchase and burn $AI tokens, creating a deflationary link to network usage. Upbit Korea listing (June 30, 2026) drove a 64% one-day rally and 1457% volume surge — direct access to Korean liquidity, traditionally a strong altcoin catalyst.

RL Swarm — paused

Per Gensyn's own documentation, RL Swarm and all Gensyn-hosted training nodes are paused, with engineering focus consolidated on Delphi. The company's framing is sequencing; the sector's reaction is that the training thesis (the project's original credibility pitch) is dormant. The most rigorous parts of Gensyn's verification stack — Verde's dispute game, Judge's evaluation primitive — were built to secure distributed training jobs. With RL Swarm paused, that verification surface has no production workload.

Tokenomics overhang

$AI trades ~$0.022 (early Aug snapshot); market cap ~$29M against fully diluted ~$220M, with ~13% of 10B supply circulating. Investor (29.6%) and team (25%) tranches together make up more than half of supply, both with a 12-month cliff from April 2026 TGE + 24-month linear unlock. Cliff ends around April 2027 — meaningful sell pressure hasn't yet hit the market.

Three watch items:

1. Whether RL Swarm resumes regular training runs (no public timeline).
2. The proposed Uniswap V3 deployment on Gensyn's L2 (raised March 2026 on Uniswap governance forum) — meant to deepen $AI liquidity ahead of the 2027 unlock cliff.
3. A completed third-party security audit of the mainnet deployment (flagged because the bridge design has been a consistently-exploited category in 2026).

Sources: Gensyn first 100 days (HOGE Wire), CMC Gensyn, docs.gensyn.ai/testnet.

Bacalhau — NCL Protocol Now Default; v1.7.0 Enterprise Hardening Lands

Bacalhau, the Compute-over-Data orchestration framework for IPFS/Filecoin/S3/edge, shipped a meaningful transport upgrade in the last fortnight:

• PR #4734 (merged): NCL (NATS Client Library) Protocol — async messaging between orchestrator and compute nodes, with separate control and data planes, ordered delivery, heartbeats, checkpoint-based recovery. Replaces the request/response bprotocol as the default in v1.6 (NCL opt-in via BACALHAU_PREFER_NCL_PROTOCOL=true); bprotocol remains the stable default for backward compat.
• PR #4687 (NCL integration tests): compute nodes now advertise their supported protocols in NodeInfo, orchestrator negotiates the preferred protocol, Buildkite runs both transport types in CI.
• Health/heartbeat config: HeartbeatInterval 15s, HeartbeatMissFactor 5, CheckpointInterval 30s, ReconnectInterval 10s base → 5m max.

The v1.7.0 release notes (covered Aug 25 but worth restating for the new transport): S3 partitioning (object / date / regex / substring distribution, automatic per-partition retries), Docker-in-Docker (DinD) compute-node image variant, and enterprise support options. Bacalhau's positioning — bring compute to data rather than move data to compute — keeps it the strongest choice for data-local analytics and regulated workloads where egress is the binding constraint.

Sources: Bacalhau v1.7.0 release notes, PR #4734 NCL Protocol, PR #4687 NCL async, bacalhau.org.

Marlin Oyster — TEE Coprocessor, x402-gateway, Verida Integration

Marlin's Oyster is the deepest TEE-first coprocessor network in DePIN compute — Intel SGX / AWS Nitro / Intel TDX + AMD SEV-SNP coverage with on-chain attestation. Three structural items worth tracking this week:

• x402-gateway (live, public repo marlinprotocol/x402-gateway): a multi-chain payment gateway (USDC on EVM and Solana) for monetizing HTTP APIs, with Oyster CVM TEE signature verification at the gateway. Pay-per-prompt AI chat with enclave-signed responses is now a packaged pattern.
• x402-ollama companion: pay-per-prompt Ollama chat with response signatures verifiable inside the CVM. A working reference for confidential inference metered by request.
• Verida integration (Verida on X): Verida's decentralized identity / storage layer uses Oyster for confidential computation over encrypted user data — agent workloads that read private user vaults without exposing plaintext to the operator.

Oyster's market position is the TEE-coprocessor layer for other networks — Verida is one customer, Bonfida's decentralized SNS gateway is another, and the OYM TEE comparison and Verifiable Compute's Hardware Problem essays are the sector's open-reference threat-model docs. The x402 stack is now the cleanest monetization layer in confidential compute.

Sources: Marlin docs, marlinprotocol/x402-gateway, Verida on X, Bonfida on Oyster.

Fluence — Spheron GPU Compute Live; Cloudless Virtual Servers in Beta

Fluence's Vision 2026 roadmap hit two milestones:

• Spheron partnership live (fluence.network/blog): GPU containers for AI workloads are available through the Fluence Platform — H200, GH200, H100, RTX5090 — at substantially lower cost than centralized cloud. Spheron is one of multiple compute providers (alongside Kabat, Piknik) supporting the network. GPU VMs and bare-metal support are coming in subsequent weeks.
• Cloudless Virtual Servers in beta (DePIN Day Singapore recap): developers can deploy workloads directly on decentralized infrastructure with transparent on-chain billing — extending the GPU compute launch to a full serverless VM model.
• FLT-backed stablecoin: users lock FLT or pFLT into a CDP vault to mint a stablecoin that becomes the preferred currency for CPU/GPU payments. Stablecoins paired with FLT in liquidity pools deepen liquidity, generate fees, and remove FLT from circulation — the first self-contained on-chain economy for decentralized compute as Fluence frames it.

The 2025–2026 AI roadmap (fluence.network) tracks: GPU node onboarding (Q3 2025 ✓), GPU container runtime support live (Q4 2025 ✓), confidential GPU computing R&D track kickoff (Q4 2025 ✓), pilot confidential job execution (Q2 2026).

Sources: Fluence GPU launch, DePIN Day Singapore recap, AI roadmap.

Golem — Quiet; Treasury-Mcap Ratio Holds

Golem remains in its dormant product phase (GLM/BTC Binance delisting, Modelserve suspension both covered Aug 11). The only substantive update is the OYM Research restated treasury position (Aug 19): holdings ~215M GLM ≈ 2.4× market cap (GLM ~$0.088, mcap ~$88M) — one of the largest treasury-to-mcap ratios in crypto. The Arkiv Network ideatcoin (Web3 databases, $4K prizes) and ecosystem-project partnerships are the only activity.

Without a Modelserve-class revival, Golem stays a balance-sheet story with a dormant product thesis. The thesis dormant until a leadership call.

Sources: OYM Research Golem, blog.golem.network.

Evidence at a Glance (Early-September 2026 Snapshot)

Freedom-Tech Takeaways & Recommendations

1. The OpenRouter provider page is now the DePIN inference storefront. Phala and AkashML are top-6 listed providers by last-week tokens. Treat the listing, model count, weekly token bar, and "Zero retention" badge as primary marketing surfaces — not blog posts. For users, the actionable sovereignty response is to confirm direct-API contracts and verify the underlying provider's chain-attestation path; the routing-layer capture is now real.

2. dstack going cross-cloud changes the confidential-compute conversation. The same workload definition now packages into Phala native, GCP TDX+TPM, and AWS Nitro EIF runtimes with one measurement-derived identity and one KMS path. The next architectural question for any team building on TEEs is no longer "which silicon" but "do I want vendor-portability (dstack), single-cloud depth (Phala native only), or a non-TEE alternative entirely." Per the OYM TEE comparison and Verifiable Compute's Hardware Problem — the silicon is the necessary root of trust, but the policy and identity layer above it is the operational question.

3. Grayscale's amended S-1 is a procedural engagement signal, not a verdict. The Aug 31 amendment keeps TAO's spot-ETF decision inside the originally-forecasted window. Plan for three scenarios: approval (institutionalizes the DeAI basket), rejection (TAO retest toward ~$164, Covenant-era decentralization question re-opened), or extension. The 43% Grayscale AI-fund TAO allocation is the manager's vote of conviction through the uncertainty.

4. Exploit Summit is the sector's single most consequential gathering in the next four weeks. New City Gas, Montreal, Sept 28–29. dTAO underwriting under ETF context, post-Covenant validator-competition restoration progress, subnet consolidation, and the inaugural 100-subnet-teams showcase. If you can only attend one DePIN event in 2026's second half, this is the one — every subnet team, validator, and institutional observer will be in the room.

5. Gensyn's verification surface needs a workload. Delphi is generating fees; RL Swarm (the original credibility pitch) is paused. The 100-day retrospective crystallizes a real risk: the most rigorous parts of Gensyn's stack (Verde, Judge) have no production workload to secure right now. Until RL Swarm resumes, the buy-and-burn economics live on Delphi's fee flow alone — meaningful, but not what the original thesis sold. Watch for any RL Swarm resumption signal before the April 2027 cliff.

6. Aethir's IDC tokenomics update is the template other DePINs may copy. Burn-mechanism + variable platform fee explicitly tied to on-chain revenue is the cleanest design statement in the sector. Watch Aethir's ATH action in the 30 days after the policy ships — if it works (revenue accrues, price stabilizes), Phala (PHA at ATL with 24B tokens/wk), io.net (IDE burns), and others have a ready-made answer to the "usage up, token down" paradox. If it doesn't, the design space narrows.

7. Gonka's governance rhythm is the credibility signal. Proposal 96 (SPRT small-sample stabilization, Aug 21) is small but well-documented. The pattern — surgical, transparent, host-action-clear proposals under audit pressure — is consistent with a team shipping the Sprint rewrite carefully. Still pre-production stability, still OTC-only, but the engineering discipline is visible.

8. Watch items: - SEC TAO spot-ETF decision — window now; watch for approval/denial/extension and the sector-beta move. - Phala cross-cloud dstack adoption — first AWS Nitro / GCP TDX enterprise customer stories; OpenRouter catalog expansion (21 → 30+ models). - Akash AEP-85 Buildpacks (Sept 30) + Annual Conference (Oct 3–5 SF) — likely venue for L1/shared-security RFP (AEP-79) news. - Aethir IDC tokenomics policy ship — burn + variable fee live on-chain. - Render Dispersed throughput on the next dashboard refresh — the single most consequential number for the Render thesis. - Gensyn RL Swarm resumption signal — any movement before April 2027 cliff. - Gonka Sprint rewrite completion — production-stability gate. - Exploit Summit (Sept 28–29 Montreal) — dTAO underwriting, ETF aftermath, subnet consolidation. - Stripe-OpenRouter integration announcements — first routing-policy changes under new ownership. - Phala throughput band — does 24B/wk hold; does dstack-on-AWS / -on-GCP land enterprise contracts.

This update synthesized from project sites and primary sources (phala.com, phala.network, akash.network/roadmap, aethir.com, exploitsummit.com, tao.media, cryptorank.io, io.net/blog, gonka.ai/docs/network-updates, primeintellect.ai, docs.gensyn.ai/testnet, bacalhau.org, marlinprotocol, fluence.network), OpenRouter provider rankings, DeAI Dashboard figures, CoinMarketCap AI analyses, CoinGecko, OYM Research, and X posts from official project accounts.