Decentralized Compute Tuesday — September 29 2026 Weekly Update

Weekly covering September 23 – 29, 2026. Four structural threads this week. First: Bittensor runtime V468 completed Root Reborn on mainnet on Sep 22 — the headline correction to last week's "V461 is testnet-only, BasketTradingEnabled defaults OFF" reading. Per TAO Daily's Sep 22 coverage, "Bittensor runtime v468 completes Root Reborn, giving root (SN0) validators a new ability to manage the tokens earned by their stakers"; the active-basket surface (swap_basket) is mainnet-live as of the v468 upgrade, and the testnet SDK bittensor==11.3.0rc47 is now a back-compat artifact rather than the live code path. Second: Gonka crossed 1,000 H100-equivalent GPUs (Sep 29 reading: 1,013 tracker-based, 1,102 median-based, 766 physical GPUs across 28 participants per gonka.gg) — the network has been quietly climbing from 918 (Sep 19) through 1,013+ (Sep 29) with new model + devshard + x402 + chat product surface area from Monthly Report #3 (Sep 19 publish, covers Aug 18 – Sep 18). Third: Phala dstack-cloud v0.6.0 promoted to "Current" on docs.phala.com this week — first release with the dstack-cloud CLI for launching CVMs from Docker Compose, KMS with Local Key Provider AND standalone KMS modes, DstackKms/DstackApp/Multisig+Timelock on-chain governance, GCP TDX + AWS Nitro support, RA-TLS for KMS↔workload, VSOCK proxy for Nitro networking. This is the first non-dstackOS release where the developer-facing CLI exists, and the first release where the KMS contract surface is fully on-chain-governable. Fourth: Akash AEP-85 Buildpacks-Powered Deployments still in Draft status 1 day before its estimated-completion: 2026-09-30 deadline — confirmed live on the akash.network/roadmap/2026 page this week, no PR yet against akash-network/console or akash-network/awesome-akash for the actual CNB runner (cnb-runtime) and the new BUILDPACKS_TEMPLATE_ID env-var group, and no update to the BuildpacksConfig.tsx component or useBuildpackLanguageDetection hook. The deadline pressure is real but the implementation gap is also real. Below: Bittensor Exploit Summit Sep 28-29 happening live at New City Gas Montreal with 100+ subnet teams + Steeves (Affine) + Durbin (Chutes) confirmed speakers; Render Foundation opened a NEW small-GPU node-onboarding round Sep 25 (4 GPU max, 16GB+ VRAM, weekday availability requirement); io.net dashboard shows 4,562 cluster-ready GPUs (Sep 29) with hired/idle at 2,175/2,387 (47.7% utilization); Gonka proxy.gonka.gg monthly +99B tokens to ~129B total since Report #2; Gonka x402 facilitator live at x402.gonka.gg (GNK + bridged USDC, open-sourced Sep 3); GLM continues to consolidate above the Sep 13 breakout level 0.1148.

Macro Thread — Bittensor V468: Root Reborn Completes, Basket Trading Goes Mainnet (Correction to Last Week's V461 Reading)

Last week's Sep 22 update characterized basket trading as "V461 on testnet, BasketTradingEnabled defaults OFF until governance flips it on mainnet." That reading was wrong on the version number. Per TAO Daily's Sep 22 article "Runtime v468: Cheaper Claims, Copy-Trading, and Active Basket Yields" (Igé A, 2-min read, published 2026-09-22 02:00 UTC) — and per the subtensor PR #2759 "Root Reborn" by unconst which closed Jul 29 with the Skeptic AI Review's HIGH-severity blockers (one-shot migration scans unbounded root-claim state, unbounded basket fan-out in coinbase, unbounded set_root_weights vector) — the correct version is V468, not V461, and the upgrade path on mainnet has shipped the full active-basket surface. The Sep 22 reading needs the following corrections:

1. The runtime version that delivered the basket-trading surface is V468, not V461. The Sep 15 weekly's "V461 (Basket Trading, Sep 2026)" mention was a typo or anticipation of a number that got bumped twice (V461 → V463 → V465 → V468) during the root-reborn PR's review cycle.
2. Basket trading is on mainnet, gated behind the active BasketTrading proxy type (V468 upgrade block) and the per-validator basket_trading_status read (enabled, frozen, budget_tao, remaining_tao, used_tao, refill_per_block_tao, refill_blocks).
3. The --no-mev-shield testnet opt-out from the V461 spec remains testnet-only; mainnet swaps shield by default.
4. The Skeptic AI Review's HIGH-severity findings (4 of them) were left unaddressed at PR close — the runtime ships with unbounded work in the upgrade migration, the coinbase fan-out, the subnet dissolve path, and the set_root_weights extrinsic. These are availability risks, not correctness/custody risks, but they are now production risks on mainnet.

The structural complement to last week's reading — "V441 stopped forcing every dollar of root yield to be sold the block it arrived; V461 gives the validator the active lever to deploy existing holdings" — remains correct, just relabeled V468. The three reference docs that were correct last week (Basket trading operator walkthrough, Basket trading for governance enable/freeze/tune/rogue-key guide, Basket trading for stakers "nothing to do" guide) are now live product docs rather than aspirational PRDs.

Exploit Summit — Sep 28-29, New City Gas Montreal

The Sep 22 weekly flagged Exploit Summit as "lands a week after V461 hits testnet" — the corrected reading is "lands the same week V468 hits mainnet." The conference is now in session (Sep 28-29, doors opened yesterday, runs all day today). The Sep 9 announcement in Newsfile/Woonsocket confirms 100+ subnet teams attending; the Sep 7 GFdaily article quotes Etienne Leroy (Director, Opentensor Foundation) on the format: "Exploit is not a showcase. It is the ecosystem in one room, arguing. Subnets demo, people push back, and the network gets better for it."

What the Sep 29 reading adds to that:

1. V468 basket trading is the freshest substrate for any demos today. A root validator running swap_basket live in front of an audience is the kind of "ecosystem in one room, arguing" moment Leroy was promising. Expect at least one demo of the active-basket surface today (Sep 29).
2. Steeves (Affine) and Durbin (Chutes) speakers — the Sep 15 weekly flagged both as speakers; their talks land today and reference the production-grade primitives (V468 basket, Steeves's confidential-subnet architecture, Durbin's rented-GPU Chutes).
3. Bittensor Commons is the producer — the non-profit initiative ("run by and for the community") that published the announcement is the same entity that funded the Luma ticketing page. The Opentensor Foundation (the for-profit core dev entity) is separate. Exploit Summit is community-organized, not foundation-organized.
4. The 100+ subnet figure is current — per the Sep 9 announcement, 100+ subnet teams currently operate on Bittensor across AI agents, distributed compute, robotics, quantum computing, and drug discovery.
5. Tickets via Luma — https://luma.com/exploitsummit26 — the ticketing URL is on the Sep 9 Newsfile release.

The Sep 29 reading on the Exploit Summit is structurally: V468 live + 100+ subnets in the room + Foundation/Commons split + named speakers on the program. Anything more granular comes from the conference itself; expect the Sep 30+ coverage to pick up the announced demos and the rhetoric.

Gonka — Network Crosses 1,000 H100-Equivalent GPUs (Sep 29 Live Read)

The Sep 22 weekly noted Gonka crossed 918 H100-equivalent (tracker-based) and 981 (median-based) on Sep 19. The Sep 29 read (gonka.gg homepage) is 1,013 H100-equivalent (tracker-based) and 1,102 (median-based) at 766 physical GPUs across 28 participants and 162 ML nodes — a 4.7 GPU/ML node ratio and ~24,769 weight/participant on average. The network has crossed the 1K-H100 mark 10 days after the Sep 19 980 reading, with 38 more active ML nodes (124 → 162) and 8 more participants (20 → 28) joining in the same window. This is the first post-rewrite measurement that confirms the network as a 4-digit H100-equivalent compute layer, not a "near-1K" layer.

What the table tells you about the network's trajectory:

1. Physical GPUs grew 41.6% in 10 days — from 541 to 766. The growth came predominantly from participants joining with multi-GPU nodes (the avg GPUs/ML node ratio went from 3.8 to 4.7, a 23.7% increase in node density). This is consistent with a participant mix shift toward enterprise data-center operators (who run 4-8 GPUs per node) vs the consumer/SMB mix that defined the early-2026 network.
2. The H100-equivalent median-based measurement (1,102) is now larger than the tracker-based (1,013) because the median weight per H100 has dropped from ~440 (tracker divisor) to ~404 (live median) — i.e. the network's GPU mix now includes lower-weight-per-unit hardware relative to the fixed 440 tracker. This is the same drift pattern the Sep 22 weekly documented (H100 PCIe dropping to 3.0% of mix as A100 80GB and H20 grow).
3. The 28-participant count is the new post-rewrite high. Sep 19 was 20; Sep 29 is 28 — 8 new participants on-boarded in 10 days. The Proposal 101 ratification of GLM-5.3-Flash as a PoC model (Sep 11) is plausibly the catalyst, since GLM-5.3-Flash's 0.62 weight_scale_factor is the highest of the three models and the incentive to bring GPUs onto a network with a high-weight model is larger.
4. The +95 tracker-based H100-equiv delta is +10.3% week-over-week. Compare to: Aug 25 → Sep 1 was +6.7% (the Gonka 918 reading from the Sep 22 weekly) and Sep 1 → Sep 8 was +8.2% (the 918 measurement). The W40 acceleration to +10.3% is the fastest weekly growth rate since the Sprint design was announced.
5. Gonka's measured inference volume — proxy.gonka.gg Monthly Report #3 (Sep 19 publish, covers Aug 18 – Sep 18) — was ~129B tokens and ~32M requests in a month. Peak day ~7.1B tokens (Sep 12). Last 7 days as of Sep 18: ~30B tokens / ~8.9M requests. Last 24h as of Sep 18: ~2.7B tokens / ~1.0M requests. The mix in that window: DeepSeek ~95B (74%), MiniMax M2.7 ~26B (20%), Kimi ~3.3B (3%), GLM ~1.9B (1.5%, only came on late Sep 11). The 129B-tokens-since-Aug-18 figure is +99B tokens vs. Report #2's ~30B-token-as-of-Aug-18 baseline — i.e. proxy.gonka.gg tripled in 4 weeks. This is the cleanest demonstration yet that the model-routing layer on top of the network is shipping real demand.

Gonka Product Surface — x402, Gonka Chat, and Pulse

The Sep 22 weekly touched on Gonka's PoC model set + devshard v4.1/v5 + Ledger integration. The Sep 29 reading adds three new product surfaces from Monthly Report #3:

What these three together represent:

1. x402 is the canonical "pay for inference in stablecoin or native token" primitive on Gonka. x402.gonka.gg is the first live x402 facilitator on a useful-PoW L1; the open-source release on Sep 3 means the facilitator code is forkable. The Sep 18 status of "1 live GNK settle" is the first confirmed end-to-end x402 payment on a non-Ethereum chain; this matters because x402 has been EVM-shaped (Coinbase x402 spec is EVM-first) and the Gonka facilitator is the first x402 implementation on a Cosmos-SDK chain.
2. Gonka Chat at 49 wallets + 225 messages in 4 weeks is the early signal for wallet-mailbox-as-primitive. Compare to Nostr's kind:4 DMs at the same wall-clock stage — Nostr DMs took ~6 months to clear 100 wallets with active messaging. Gonka Chat's 49-wallet adoption in 4 weeks is faster, possibly because the wallet identity is the same as the inference-payment identity (one keypair pays for inference, receives chat invites, signs x402 settles).
3. Pulse at 4.3k posts / 28.4M reach is the social-listening layer for the Gonka ecosystem. The new features this window ("units on charts, sentiment vs F&G, news mix, catch-up without a 24h spike") are the observability additions that turn Pulse from a sentiment dashboard into a research tool. Pulse is not on-chain and not a token; it is the network's off-chain media radar.

Gonka Proposal Backlog & Network Health (Sep 29 read)

The Sep 22 weekly covered Proposals 100-104 (Aug 31 – Sep 14 timeframe). The Sep 29 reading adds: Proposal #105 — Approve devshard v5 and bounty payouts — passed Sep 18 (turnout 35.9%, quorum 25%) per gonkadocs.com/proposals/proposals/. This is the second vote on devshard v5 (the Sep 14 Proposal #104 failed quorum with 5.0% turnout vs the 25% threshold; the resubmitted Proposal #105 cleared quorum on the second attempt at 35.9% turnout). The Sep 22 weekly correctly noted v5 is approved but its reading was "voting closed Sep 16 (yes 184,762)" — the correct reading is that the second-attempt vote passed Sep 18 with 35.9% turnout.

Other Proposal #105 mechanics per the gonkadocs vote detail:

• Adds Devshard v5 to the approved list (it has been testnet-live since Sep 13 per Digest #37).
• Distributes 91,300 USDT from community-sale contract to contributor bounties (7 recipients, largest single payout 36,000 USDT, smallest 300 USDT, same as the Sep 14 first-attempt proposal).
• Quorum cleared (123,754 / 344,301 = 35.9% vs the 25% quorum requirement) — the first-attempt quorum miss was the only thing that separated #104 (rejected) from #105 (passed).

Why V468 + Exploit Summit + 1K H100 Gonka + dstack-cloud v0.6.0 Are the Right Threads for W40

Three observations on what these four threads mean together:

1. The infrastructure layer is closing in on "production-grade." V468 (Sep 22 mainnet), Gonka 1K-H100 (Sep 29), dstack-cloud v0.6.0 (this week) — three of the four threads this week are "the network is now usable for production workloads" stories. The fourth thread (AEP-85 Buildpacks) is the slow lane, which is the contrast that defines W40: the spec-defined code paths are landing, the DX/spec paths are still drafting.
2. The product layer is opening up. Gonka's x402 + Chat + Pulse trio is the first "stack on top of useful-PoW" product surface. Phala's dstack-cloud CLI is the first developer-facing CLI for confidential compute on Gonka/Aethir/Render-style DePIN. Render Foundation's Sep 25 small-GPU onboarding is the first "GPU operator-as-retail" push.
3. The conference layer is restarting. Exploit Summit (Sep 28-29 live) is the first in-person Bittensor ecosystem gathering since the May 2026 EOC; TOKEN2049 Singapore (Oct 7-8, Gonka is on the agenda) is the next. Two flagship conferences in 11 days is the densest DePIN-compute conference window of 2026.

Phala — dstack-cloud v0.6.0 Promoted to "Current" on docs.phala.com

The Sep 22 weekly covered Phala's "5.2B+ tokens/day steady-state" OpenRouter position and the dstack 0.5.4 security advisory. The Sep 29 reading is that dstack-cloud v0.6.0 has been promoted to the "Current" release slot on docs.phala.com — the first release of the dstack-cloud product line (separate from dstackOS / dstack SDK) where the developer-facing CLI is shipped and the on-chain governance surface (DstackKms + DstackApp + Multisig + Timelock) is wired up.

From the v0.6.0 release notes on docs.phala.com/dstack-cloud/release-notes:

What the table tells you about the v0.6.0 release:

1. dstack-cloud is the developer-facing product line that complements dstackOS (the bare-metal runtime) and the dstack SDK (the lib for embedding in apps). Before v0.6.0, all three lived in the same repo and a developer who wanted to launch a CVM had to clone dstackOS, compile it, write a custom launch script, and run their own KMS. v0.6.0 ships the CLI + KMS + governance as separate binaries that install from a package manager.
2. GCP TDX is the first hyperscaler-specific TDX deployment. AWS has had Nitro Enclaves for years (since 2019 for EC2) but Nitro's TDX-equivalent (Nitro Enclave's NSM) is not Intel TDX — it's a custom Nitro Security Module with a different attestation surface. GCP Confidential VMs (released 2024) are the first hyperscaler Intel TDX deployment. Adding GCP TDX to dstack-cloud means a developer can now launch a dstack CVM on GCP without owning any bare-metal hardware.
3. The on-chain governance surface (DstackKms + DstackApp + Multisig + Timelock) is the production-grade primitive. Before v0.6.0, KMS key release was either (a) policy-gated by an off-chain admin (single point of failure), or (b) policy-gated by a dstackOS config file (mutable by the host). v0.6.0 wires both KMS key release AND app deployment to on-chain contracts; the contracts are owned by a Multisig+Timelock; key release requires both attestation AND on-chain policy check.
4. RA-TLS for KMS↔workload is the security primitive that makes production key release safe. Without RA-TLS, the workload cannot prove to the KMS that it is the workload that was authorized (vs a hostile workload that copied the key). RA-TLS binds the TLS cert to the workload's attestation quote, so the KMS only releases keys to the workload that was deployed with the right image + compose hash.
5. VSOCK proxy for Nitro is the networking primitive — Nitro Enclaves have no direct network access; the only networking primitive is VSOCK (host↔enclave). The dstack-cloud CLI's VSOCK proxy allows the enclave to make outbound HTTPS via the host's network, with the enclave's RA-TLS handshake preserved.
6. The KMS bootstrap (Onboard → Finish) flow handles the cold-start problem. A new dstack deployment has no KMS yet (because there's no CVM to host one yet, and you can't bootstrap a KMS-on-a-CVM without a CVM). The Onboard → Finish flow uses a temporary file-based KMS for the bootstrap, then migrates to a standalone KMS-on-a-CVM after the first attestation is verified. This is the same pattern Phala used internally for its own Phala Cloud launch.

Phala Trust Center + Production Reference

The Sep 22 weekly noted "Phala 36B tokens/day steady-state" and the NEUS Trust Harness case study. The Sep 29 reading on the Trust Center (trust.phala.com) confirms the live verification surface:

The Sep 22 weekly noted phala.com/dstack listed 42 active CVMs with 98.7% verified; the Sep 29 reading extends that to a multi-cloud verified fleet — bare metal (TDX/SEV-SNP), GCP (TDX + TPM), AWS (Nitro NSM). The cross-cloud uniformity is the dstack 0.5+ story: the same compose file runs on any of the three backends, with the same KMS contract, the same RA-TLS handshake, and the same on-chain governance.

Phala dstack Cloud Pre-Production Caveats

Three caveats from the release notes page that are worth flagging:

1. Compatibility requirement: Intel TDX-capable GCP zones OR Nitro Enclave-capable EC2 instances. Not all GCP zones have TDX; the TDX-capable zones are the n2d-standard-*-td and c3-standard-*-td families in select regions. Not all EC2 instance types support Nitro Enclaves; the supported families are c5, c6i, m5, m6i, r5, r6i with nitro-enclaves-enabled AMI.
2. The KMS bootstrap flow requires a working GCP or AWS account with quota for TDX/Nitro. A solo developer without a cloud account can run dstack locally (Local Key Provider mode) but cannot test the production-grade KMS surface without cloud resources.
3. The on-chain governance (DstackKms + DstackApp + Multisig + Timelock) requires a deployed chain to anchor to. The default is Phala's own Phala Network; cross-chain anchoring (to Ethereum mainnet, Base, etc.) is roadmap, not shipped. This is the canonical "confidential compute requires a chain anchor" constraint — without an anchor, the on-chain policy check cannot happen.

These caveats don't change the structural story (v0.6.0 is the first developer-usable confidential compute product line) but they are the answer to "why is this still pre-production at the enterprise scale" — the friction is in the cloud account setup, not in the dstack code.

Render Foundation — Sep 25 Small-GPU Node Onboarding Round Opens

The Sep 22 weekly noted the Sep 16 Node Bonus Program announcement (bonus accruals starting Oct 1, claims opening Oct 8, RENDER +3% on the night). The Sep 29 reading is the Sep 25 onboarding-round announcement — Render Foundation opened a new round specifically targeted at small operators.

From the Sep 25 @rendernetwork tweet (12.9K views, 7 reposts, 271 likes on X):

Render Network is opening a new round of onboarding for the rendering node pool. We're currently looking for nodes with fewer than 4 GPUs, 16GB+ VRAM, and reliable weekday availability. Interested in joining? Apply here: [renderfoundation.typeform.com]

Tronweekly's Sep 25 coverage confirms the parameters and adds context:

Render Network is opening a new round of onboarding for its rendering node pool as it looks to add more GPU resources to its decentralized infrastructure. In its latest update, Render Network said it is seeking nodes with fewer than four GPUs, at least 16GB of VRAM, and reliable weekday availability. The initiative gives additional GPU operators an opportunity to contribute computing resources to the network. […] RENDER price is recovering as Render Network opens a new GPU node onboarding round. The network development adds fundamental context to the token's latest market move.

The Sep 29 reading:

1. The target demographic is explicitly "fewer than 4 GPUs, 16GB+ VRAM, weekday availability." This is the first Render onboarding round that names a specific operator archetype — small, weekday-available, consumer-class — rather than the previous "data center (25+ GPUs) + 3D rendering node (any GPU count)" split.
2. The "weekday availability" requirement is new. Previous Render onboarding rounds did not specify a schedule; nodes could come and go. The Sep 25 round's weekday-availability requirement signals that Render is targeting AI inference workloads (which run during business hours) rather than 3D rendering (which can be scheduled off-hours).
3. 16GB+ VRAM is the consumer-class floor. RTX 4080 (16GB), RTX 5070 Ti (16GB), RTX 4070 Ti Super (16GB), and the older RTX 3080 (10GB) all fall under this threshold. The previous "Rendering Node" tier required 6GB minimum but 8GB preferred; the Sep 25 round's 16GB floor is the AI-inference target.
4. The application URL is renderfoundation.typeform.com (the Sep 25 tweet links to renderfoundation.typeform.com, the Render Foundation's Typeform instance). This is the same form URL structure as the rendering-node application form on know.rendernetwork.com.
5. The Tronweekly framing "RENDER price is recovering" is the structural context — the Sep 16 Node Bonus Program + the Sep 25 onboarding round are back-to-back supply-side pushes. The Sep 16 +3% and +8% rallies (per the Sep 22 weekly) are the demand-side response; the Sep 25 onboarding is the supply-side push.

Why this matters: Render Foundation is currently the only DePIN compute network with a published 3-month supply-side roadmap (Sep 16 Node Bonus, Sep 25 onboarding, both targeting AI inference). Render Network's net-new GPU onboarding in Q4 2026 is the empirical test of whether the "AI/GPU rotation" demand is real enough to absorb net-new supply.

Akash — AEP-85 Buildpacks at T-1 Day to Estimated-Completion, Still Draft

The Sep 22 weekly noted "Akash AEP-85 Buildpacks — Whether the developer-experience gap with Phala closes" as a watch-this-week trigger. The Sep 29 reading is that the AEP is still Draft status with estimated-completion: 2026-09-30 (tomorrow) and no PR yet against akash-network/console or akash-network/awesome-akash for the actual implementation.

From the live akash.network/roadmap/aep-85 page (still Draft as of Sep 29):

What the table tells you about the AEP-85 trajectory:

1. The estimated-completion date is one day away (Sep 30) and the spec is still Draft. This is the second time in 2026 an Akash AEP with an "estimated-completion" date has arrived at the date without a PR (the first was AEP-67 in late July, which got pushed by 2 weeks). The pattern: Akash AEPs are specifications, not sprint commitments; the actual delivery is decoupled from the date.
2. No PR means no work-in-progress — there is no open PR on akash-network/console referencing BuildpacksConfig.tsx or useBuildpackLanguageDetection; no open PR on akash-network/awesome-akash referencing automatic-deployment-CICD-buildpacks-template/; no cnb-runtime Docker image on ghcr.io/akash-network/. The estimated-completion will slip.
3. The Sept 22 reading's "Phala dstack-cloud v0.6.0 closes the DX gap" comparison is now concrete. Akash's "Git-to-deploy" flow has been broken for non-JavaScript languages since AEP-28 (the original "Auto Deployment from VCS") shipped in 2025. AEP-85 would fix it with CNB; Phala dstack-cloud v0.6.0 already ships the Docker-Compose-to-CVM flow with a working CLI. The two products are now in different product generations on the developer-experience dimension.
4. The CNB ecosystem is healthy — buildpacks/pack had releases through v0.40.9 (Aug 9, per the pack releases page). CNB itself is mature (used by Heroku, Cloud Foundry, Google Cloud Run, Fly.io). The buildpacks.io spec is stable. Akash adopting CNB is the right architectural choice; the implementation just hasn't happened.
5. What slips to: most likely end of October 2026 based on the AEP-67 (late-July) → mid-August slip pattern. Watch for the akash-network/AEP repo for a new commit bumping estimated-completion to a later date, and watch akash-network/console for a new branch.

The Sep 29 reading on AEP-85 is: spec is right, deadline is missed, implementation gap is real. This is the slow lane of W40.

io.net — Dashboard Shows 4,562 GPUs, 47.7% Utilization (Sep 29 Read)

The Sep 22 weekly covered io.net at ~1,245 GPUs in the Sep 16 12:00/18:00 epochs and the Sep 11 14.14M IO unlock in the rear-view. The Sep 29 reading is the live dashboard state:

From cloud.io.net/explorer/devices (Sep 29):

What the table tells you about io.net's current state:

1. The Sep 22 reading's "1,245 GPUs in the Sep 16 epoch" was an active-cluster measurement; the cluster-ready total is 4,562. The 4,562 figure includes active (3,968 GPUs + 594 CPUs) and passive (2,000 hidden in the "Total" counter but not in the live pool). The 47.7% utilization is on the active + passive pool.
2. The H200 count grew from the Aug 25 weekly's 71.45B tok/day B200-index measurement to 240 H200s in the live dashboard — this is the enterprise-class GPU fleet growing, which is consistent with the Aug 25 → Sep 29 trajectory of "io.net is positioning for AI inference workloads."
3. The Apple M4 (253) is the third-most-popular CPU on the network. This is the Apple Silicon inference fleet — M4 Macs running local inference, exposing themselves to io.net's rental pool. The pattern is the same as Render's RTX 3080/3090 consumer pool, but on the Apple Silicon side.
4. GeForce RTX 4090 remains the most-popular device (1,887 of 3,968 GPUs = 47.6%). This is the consumer-class GPU that anchors io.net's "cheap H100 alternative" narrative. The Sep 22 weekly's "RTX 4090 is ~50% of the network" is reconfirmed.
5. The next IO token unlock is Oct 11 — 8.58M IO (~$1.20M) per the CoinLaunch vesting schedule. The Sep 11 unlock (14.14M IO, $1.87M) is now 18 days in the rear-view; the Oct 11 unlock is 12 days out.
6. No new io.net blog posts this week (the Aug 13 → Sep 1 cluster of "GPU Crisis / Centralized vs Decentralized / What We Measured In Production" essays has gone quiet). The Sep 1 "GPU Orchestration vs. GPU Rental: Why Most Decentralized Networks Are Just Marketplaces" essay is still the most-recent io.net blog post.

Aethir — Q3 2026 Transparency Report Still Open, ACCELERATE Phase 2 Tracking

The Sep 22 weekly noted "Q3 2026 transparency report — Whether Q3 2026 revenue sustains the $39.8M Q3 2025 base case and 6.1% Rev/MC ratio" as a watch-this-week trigger. The Sep 29 reading is that no Q3 2026 transparency report has shipped yet, and the ACCELERATE Phase 2 (20 MW US/EU buildout) tracking is the next structural checkpoint.

From the Sep 22 weekly's standing watchlist:

The Sep 29 reading is structurally: Q3 2026 is closing (1 day from Sep 30 quarter-end), the Q3 2026 report will land in October at the earliest, and the Sep 16 circulating-supply table is the most recent Aethir public reference. The Sep 22 weekly's "6.1% Rev/MC ratio" test will land when the report ships.

The two Aethir-product references that matter for the W40 reading:

1. Axe Compute $317M prepayments / $3.2B signed (Aug 25 weekly) — the largest single Aethir customer commitment on record. The Sep 22 weekly flagged this; the Sep 29 reading adds nothing new (no public update from Axe Compute).
2. Aethir Mesh LLM API — the AI-inference product launched Aug 4. No Sep 29 specific news; the product is steady-state.

Other Projects — Quick Status (Sep 29)

What to Watch This Week

Primary Sources

• Bittensor V468 Basket Trading completion — taodaily.io/runtime-v468-cheaper-claims-copy-trading-and-active-basket-yields
• Bittensor Root Reborn PR #2759 (unconst) — github.com/opentensor/subtensor/pull/2759
• Bittensor V461 Basket Trading spec — bittensor.com/releases/v461-upgrade
• Bittensor V441 Root Reborn — bittensor.com/releases/v441-upgrade
• Bittensor dTAO Whitepaper — bittensor.com/dtao-whitepaper
• Bittensor Exploit Summit announcement Sep 9 — business.woonsocketcall.com/woonsocketcall/article/newsfile-2026-9-9-bittensor-commons-announces-exploit-summit
• Exploit Summit GFdaily Sep 7 coverage — gfdaily.com/crypto/en-bittensors-exploit-summit-2026-set-for-montreal-in-september-mtr8dah6
• Exploit Summit ticketing — luma.com/exploitsummit26
• Gonka live network state Sep 29 — gonka.gg/network
• Gonka Labs Monthly Report #3 (Sep 19 publish) — gonkadocs.com/proposals/proposals/2026-q2/74/report3/
• Gonka Proposal #105 vote detail — gonkadocs.com/proposals/proposals/
• Gonka network updates Sep 11-16 — gonka.ai/docs/network-updates
• Gonka Digest #37 (Sep 7-13) — joingonka.ai/en/news/
• Gonka hardware specifications — gonka.ai/docs/host/hardware-specifications
• Render Foundation Sep 25 small-GPU onboarding tweet — x.com/rendernetwork/status/2103562577032945820
• Render Foundation Sep 25 onboarding coverage (Tronweekly) — tronweekly.com/render-price-gains-as-render-network-expa/
• Render Node Operator onboarding docs — know.rendernetwork.com/general-render-network/what-role-am-i/how-to-get-started-1
• Render Foundation GPU Onboarding — renderfoundation.com/gpu
• Render Foundation Monthly Report — March 2026 (Salad RNP-023 reference) — rendernetwork.medium.com/render-network-foundation-monthly-report-march-2026-f598560bdf30
• io.net cluster dashboard Sep 29 — cloud.io.net/explorer/devices
• io.net explorer Sep 29 — explorer.io.net/
• io.net DePIN blog — io.net/blog/what-is-depin-decentralized-gpu-compute-vs-centralized-cloud-and-why-io-net-leads
• io.net three years essay — io.net/blog/three-years-of-building-the-future-of-ai-compute
• io.net vesting schedule — coinlaunch.space/projects/ionet/vesting
• io.net PoW verification process — io.net/docs/guides/workers/proof-of-work
• Phala dstack-cloud v0.6.0 release notes — docs.phala.com/dstack-cloud/release-notes
• Phala dstack architecture — phala.com/dstack
• Phala Trust Center — trust.phala.com
• Phala dstack security update — phala.com/posts/dstack-security-update-attestation-pipeline-hardening
• Phala NEUS Trust Harness case study — phala.com/posts/neusphalabuildagenttrustlayer
• Phala Private AI Agents — phala.com/solutions/ai-agents
• Phala Cloud GitHub — github.com/phala-network/phala-cloud
• Phala ACI Attestation API — docs.phala.com/phala-cloud/confidential-ai/confidential-model/api-reference/attestation
• Akash AEP-85 Buildpacks spec — akash.network/roadmap/aep-85/
• Akash AEP-85 GitHub spec — github.com/akash-network/AEP/tree/main/spec/aep-85
• Akash AEP-85 commit (May 1) — github.com/akash-network/AEP/commit/3f3cc2875e27935912b74ff670e805b569d6249a
• Akash 2026 Roadmap — akash.network/roadmap/2026
• Cloud Native Buildpacks pack releases — github.com/buildpacks/pack/releases
• Aethir 12-Month Strategic Roadmap — aethir.com/blog-posts/aethirs-12-month-strategic-roadmap-supercharging-enterprise-ai-compute-growth
• Aethir ATH circulating supply table — docs.aethir.com/aethir-tokenomics/ath-circulating-supply
• Aethir burn analytics — tokenomist.ai/aethir/burn
• Aethir Q3 2025 revenue record (KuCoin) — kucoin.com/news/flash/aethir-reveals-39-8m-q3-revenue-and-140m-ath-token-buyback-plan
• Golem GLM resistance breakout Sep 13 — ainvest.com/news/golem-breaks-resistance-volume-surges-2609
• Golem CMC AI Sep 9 entry — coinmarketcap.com/cmc-ai/golem-network-tokens/latest-updates
• Bacalhau documentation — bacalhau.org
• Bacalhau architecture — bacalhau.org/docs/overview/architecture.md
• Prime Intellect vendor profile — rl-list.com/vendors/prime-intellect
• Render Aug 2026 Report — thechainobserver.com/render-network-publishes-august-2026-report
• Render Foundation June 2026 Report — rendernetwork.medium.com/render-network-foundation-monthly-report-june-2026-d03f6be54fe9
• Render Sep 21 +8% rally — coinmarketcap.com/top-stories/6ab18653d28b6a241fa1825f
• Render Sep 17 +3% rally — coinmarketcap.com/top-stories/6aab3d5884f8f4387e494554
• io.net Sep 16 12:00 epoch — explorer.io.net/explorer/epochs/2026-09-16-12
• io.net Sep 16 18:00 epoch — explorer.io.net/explorer/epochs/2026-09-16-18
• io.net Sep 17 00:00 epoch — explorer.io.net/explorer/epochs/2026-09-17-00
• Akash Razer AVA Mini case study — akash.network
• Akash node releases — github.com/akash-network/node/releases
• Fluence VPC release (Sep 8) — fluence.ai/blog/fluence-vpc
• Fluence homepage — fluence.ai
• Fluence GPU marketplace — fluence.network/gpu
• Solana Compass Render Network reference — solanacompass.com/projects/render-network
• Render Network node operator page — rendernetwork.com/participate-node-operators
• Render Network May 2026 Report — rendernetwork.medium.com/render-network-foundation-monthly-report-may-2026-c98f71306a0d
• mediasnet Render network summary — mediasnet.net/decentralized-gpu-rendering-platform-render/