Decentralized Compute Tuesday — September 22 2026 Weekly Update

Weekly covering September 16 – 22, 2026. Five structural threads this week. First: Bittensor's V461 Basket Trading spec hit testnet on the week of Sep 17 — the runtime upgrade that hands root-validator yield allocation to active validators (away from the set_root_weights dividend-deployment vector that V441 Root Reborn had already replaced for non-fund holders) and that gates swap_basket behind BasketTradingEnabled until governance flips it on; 11.3.0rc47 is the first fully-working testnet SDK after a broken 11.3.0rc46. Second: Gonka's network crossed 980 H100-equivalent GPUs (median-based) on Sep 19 at 20 active participants — the first post-rewrite measurement that puts the network near the 1K-H100 mark the team has talked about since the Sprint design — with GLM-5.3-Flash ratified as a PoC model on Sep 11 (Proposal 101) and the devshard v4.1/v5 stack now running alongside the older v3. Third: Render's Burn Tracker shows ~27.97% of daily emissions burned on Sep 16 (4,588 RENDER across 183 transactions) — and RENDER climbed 3% in the same evening window as the burn readout hit and +8% by Sep 21 on the AI/GPU rotation — making the burn fraction, not the burn dollar amount, the new headline metric for RENDER (the dollar notional is small but the percentage of supply removed is high). Fourth: io.net settled at ~1,245 GPUs in the Sep 16 12:00 / 18:00 epochs (down from the 71.45B tok/day B200-index peak of the Aug 25 weekly) — and the Sep 11 14.14M IO unlock ($1.87M) is now in the rear-view, with the next unlock on Oct 11 (8.58M IO, ~$1.20M) per CoinLaunch's vesting schedule. Fifth: Fluence shipped VPC on Sep 8 (the first major "non-GPU" infrastructure release of 2026) — VMs now take custom cloud-init, isolated private subnets, public IP whitelisting, and security groups, with Tailscale drop-in as the canonical use-case for the cloud-init path. Below: Akash Razer AVA Mini case study (15x cheaper than centralized APIs, $0.01/image), Aethir ATH circulating-supply Sep 16 read (23.31B ATH = ~9.0% of 42B cap), Phala dstack 0.5.4 security advisory (LUKS header validation CVE) as the canonical security anchor, GLM breaks resistance Sep 13 (857K volume, +17.36% week).

Macro Thread — Bittensor V461 Basket Trading: Root Validators Get Active Rebalancing, swap_basket Gated Off

Bittensor's V441 (Root Reborn, July 2026) had already moved root-yield deployment away from the old set_root_weights vector and onto validator-driven alpha allocation. V461 (Basket Trading, Sep 2026) closes the second half of that transition by giving root validators an explicit swap_basket extrinsic to actively rebalance the holdings of their root-managed fund — and removes the last vestige of the set_root_weights machinery (call 146, the RootWeightSettingEnabled gate, the RootWeightsCap map, the validator_root_weights read API, and the weights field of BasketSummary). The Sep 22 reading is the spec itself, the testnet SDK, and what is/isn't on mainnet.

What the table tells you about the design:

1. The dividend-vs-trading split is now explicit. V461's release notes say it bluntly: "Until now a fund could also be steered through the dividend stream: a set_root_weights vector decided where new yield was deployed. That design is removed — dividends now always accumulate in place on the subnet they were earned on." This is the structural complement to V441's "we hand that stream to the people best positioned to manage it" line: 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, with a MEV-shielded mainnet wrapper and a testnet-unshielded trial.
2. Netuid 0 is now a first-class holding. "Netuid 0 on either side is the fund's TAO cash slot, so a validator can also move into or out of cash." This is the missing piece that makes a root-managed fund function like an actual index fund — without cash, you can only rebalance within alpha holdings.
3. Mainnet is gated behind governance. BasketTradingEnabled defaults OFF, and swap_basket "fails with BasketTradingDisabled until governance flips BasketTradingEnabled on." This is a deliberately conservative launch — V461 is the SDK and the spec; mainnet activation is the next on-chain vote. The Sep 22 reading is that the upgrade is ready to be flipped on by token holders; it is not yet on mainnet.
4. The SDK surface is small and intentional. btcli root swap --from --to --amount [--max-slippage] [--hotkey] [--proxy-for] quotes both legs, sets the floor to the quote less --max-slippage percent (default 1%), and shows "origin and destination holdings, the expected and minimum output, and how much of the bucket is left before you sign." The "bucket" reference is the budget_tao mechanism that prevents a single validator from draining the fund via rapid swaps; on mainnet, swaps are MEV-shielded by default (--no-mev-shield is the testnet-only opt-out because the shield's validator-side reveal runs on mainnet only).
5. No validator's income moves. "Weights only ever decided how an earned dividend was deployed, never what a validator earned, so no validator's income moves." This is the key reassurance — the migration of stored root vectors is purely operational, not economic.
6. The Basket trading operator walkthrough, the Basket trading for governance guide (enable, freeze, tune, rogue-key playbook), and the Basket trading for stakers guide ("nothing to do, and how to watch") are the three reference documents. The governance guide is the one that matters for the next on-chain vote: it lays out the procedure to flip BasketTradingEnabled on, freeze it, tune the budget, and handle the rogue-key scenario (a validator that tries to drain the fund).

Why V461 Matters for the Wider Subnet Economy

Three observations:

1. The basket is the first primitive that lets a TAO holder delegate capital-allocation decisions to a validator they trust, without giving that validator custody. A TAO holder stakes to root, a validator is chosen, the validator runs swap_basket, and the resulting alpha composition is on-chain. The TAO holder can withdraw their stake at any time and exit at the current composition. This is structurally different from staking to a subnet directly (which exposes the holder to that subnet's alpha price and that subnet's emission schedule) and from holding TAO (which gets the base 6.3%/yr yield but no upside on specific subnets).
2. The MEV shield on mainnet is the production-grade primitive. "On mainnet leave that flag off; swaps shield by default." The shield uses a validator-side reveal to prevent front-running; the testnet-only opt-out exists because the reveal protocol runs on mainnet validators, not testnet ones. This is the same MEV-shielded pattern that Bittensor used for commit_reveal_vote weights; V461 is the first time it's applied to a value-moving extrinsic.
3. The Sep 28-29 Exploit Summit speaker list (from the Sep 15 weekly) gains a new relevance with V461. Jacob Steeves (Affine CEO) and Jon Durbin (Chutes/SN64) are both on the speaker list; V461's basket-trading primitive is a natural reference for their talks on confidential-subnet infrastructure (Steeves) and production subnet inference (Durbin). The Sep 22 reading is that the Exploit Summit will land a week after V461 hits testnet, which means the room will have a fresh spec to argue about.

Gonka — Network Crossed 980 H100-Equivalent GPUs (Sep 19), GLM-5.3-Flash Ratified Sep 11

The Sep 15 weekly noted that the Gonka post-rewrite measurement had crossed 918 H100-equivalent GPUs (tracker-based, ÷440) and 981 (median-based). The Sep 19 reading (per gonka.gg/network) shows 918 H100-equivalent (tracker-based) and 981 (median-based) at 541 physical GPUs across 20 active participants and 142 ML nodes — i.e. a 3.8 GPUs/ML node ratio and 20.2K weight/participant on average. This is the first post-rewrite measurement that puts the network inside 2% of the 1K-H100 mark the team has referenced since the Sprint design was announced. The Sep 22 reading is the live network state plus three Sep 11-16 governance actions that define the model set.

What the table tells you about Gonka's trajectory:

1. The PoC model set is now: MiniMaxAI/MiniMax-M2.7 (weight_scale_factor 0.3024, penalty_start_epoch 278) + deepseek-ai/DeepSeek-V4-Flash-0731 (0.246, epoch 360) + zai-org/GLM-5.3-Flash (0.62, epoch 394). The GLM-5.3-Flash weight (0.62) is the highest of the three models and is what drives the network's measured inference volume. The activation at epoch 394 means the non-participation penalty (15%/epoch against full weight) applies for any host that hasn't declared intent; this is enforced by the penalty_start_epoch parameter.
2. Devshard v5 is the first non-trivial runtime change since the rewrite. Per the Sep 14 proposal: "v5 is the first update in a series of devshard and network-node changes that refactor the monolith toward a high-availability, fault-tolerant, scalable architecture." The v4.1 (passed Sep 11) is the compatibility shim; v5 (passed Sep 16) introduces height-sync — every host and the sequencer keep a signed, replayable view of mainnet height inside the escrow log, including when there is no inference traffic — and diff heightstamping (observed_height / observed_block_hash stamps on all messages that previously carried timestamp fields).
3. The 91,300 USDT bounty payout (Sep 14 Proposal 104) is split across 7 recipients. The largest single payout is 36,000 USDT (gonka1ejkupq3cy6p8xd64ew2wlzveml86ckpzn9dl56); the smallest is 300 USDT. This is the contributor-bounty loop the network has used since the rewrite — the same pattern that funded the kaitaku.ai team's GLM-5.3-Flash proposal (which was independently validated by vbgd0 before being jointly proposed).
4. The Ledger integration (Quant Mesh Limited, $350,350 USDT) is the first on-chain ratification of a non-crypto vendor. Per the Sep 11 announcement: "Quant Mesh Limited is ratified as the legal contracting counterparty for the Gonka-Ledger integration, and 350,350 USDT has been allocated from the Community Pool to gonka1yqj5xf0wtqgpdmv5v68cus0tp2j5fv7lzcfd6g." The breakdown: 275,000 USDT Phase 1 Launch Fees + 75,000 USDT first-year Annual Activation Fee + 350 USDT processing costs. The Sep 22 reading is that Gonka is building out the hardware-wallet integration path that would let inference payments settle to a Ledger device — which is the first credible cold-storage story for any DePIN compute network.
5. The model-args for GLM-5.3-Flash are now public: --max-model-len 400000 --kv-cache-dtype fp8 --enable-auto-tool-choice --tool-call-parser glm47 --reasoning-parser glm45 --trust-remote-code. The 400K context window is the largest of any PoC model on Gonka; the fp8 KV cache is the memory-optimization that lets the model fit on H100/H200 80GB. The glm47 tool-call-parser and glm45 reasoning-parser are the upstream Z.ai conventions; Gonka doesn't fork them, it adopts them.

Gonka Live Network State (Sep 19 read)

Two observations:

1. H100 PCIe has dropped out of the top half of the mix. The Sep 19 read shows 16 H100 PCIe (3.0%) — the same as B200 (16, 3.0%). This is structurally different from io.net's mix (where RTX 4090 is ~50%) and from Aethir's mix (where H100/B200 dominates because of the enterprise customer base). Gonka's mix is the most enterprise-grade of any DePIN compute network on a per-GPU basis — the median GPU is H100 80GB HBM3 or H200, not consumer silicon.
2. The Sept 19 10.33B tokens/24h figure (per power.gnk.space legacy dashboard) is a 26.7% drop from the prior day, with completion rate at 45.6% (up 19.4pp). The drop is partly explained by the GLM-5.3-Flash activation penalty: any host that didn't declare intent by epoch 394 lost 15% of their weight on Sep 16-17, which directly reduces the network's inference capacity. The Sep 22 reading is that the network is still finding its new equilibrium under the v5 devshard stack.

Render — Burn Rate as the Headline Metric, +8% on AI/GPU Rotation

The Sep 15 weekly flagged that Render's burn rate is a more useful metric than burn dollar amount — a 279% YoY increase in token burns (per HOGE Wire's Sep 3 coverage) at a price near $1.43 means more tokens are burned per day than the year-ago comparable. The Sep 16-21 reading is the burn-tracker data and the price action. On Sep 16, the dedicated Render burn tracker reported 4,588 RENDER (~$6,212) burned across 183 transactions, corresponding to ~27.97% of that day's emissions removed from circulation (per the burns-update tweet captured by CMC AI). In the same evening window RENDER moved from ~$1.29 to ~$1.33 (+3.10% over several hours).

What the table tells you about Render's current state:

1. The burn ratio (27.97%) is the new headline metric. RENDER's daily emissions schedule runs ~9.1M tokens in year one (per HOGE Wire's Sep 3 coverage) declining to ~5.9M; if 27.97% of those emissions are burned the same day, net daily issuance is negative by ~16.7% of gross emissions. This is the structural argument that RENDER has flipped from net-inflationary to net-deflationary at the protocol level — and the Sep 16 burn readout is the first public confirmation that the percentage, not the dollar notional, is the metric that matters.
2. The Node Bonus Program (announced Sep 16 8:47pm UTC) is a targeted supply-side incentive. Per the Render infrastructure account: "rendering node operators receive an extra RENDER allocation based on past frames rendered, with bonus accruals starting Oct 1 and claims opening Oct 8." This is the network's response to the supply-constrained environment flagged in the Aug 25 weekly — Render hit its first negative GPU supply event in mid-2026, and the bonus program is an explicit mechanism to bring more consumer GPUs onto the network.
3. The Sep 21 +8.45% move is mostly market beta, not Render-specific. The total crypto mcap move of +5.1% and altcoin mcap of +3.3% explain most of the directional move; Render's outperformance is +3.45% above the altcoin basket. The technical breakout framing (descending channel + falling wedge) is the proximate cause of the move but the underlying demand was the AI-coin rotation.
4. The Aug 25 weekly's "Render spent 2026 supply-constrained and burning tokens, yet RENDER still sits near $1.43" tension is now resolved differently. RENDER is at ~$1.79 (Sep 21 high) — still 87% below the March 2024 ATH of $13.53 — but the tokenomics story has converged: 27.97% daily burns, Dispersed AI subnet (95% burn rate per Messari), Dispersed + Salad subnet integration (RNP-023), Octane 2027.1 neural-rendering mode (July 2026), RENDER as a payment method in OTOY's Studio for 30+ generative models. The Sep 22 reading is that the burn fraction is now the metric the network is being valued on, not the dollar notional.
5. The Salad subnet integration (RNP-023) is "completed: RENDER payments — shipped" plus "in progress: Chef rewards in RENDER (testing)" per the June 2026 Render Network Foundation Monthly Report (carried in the Sep 22 reading because the September 2026 report has not yet been published). The full subnet integration is "coming soon" per the report; the Sep 22 reading is that the RNP-023 workstream is the primary structural path for adding the 60,000 Salad GPUs to the Render network, paid in RENDER with revenue driving burn.

The Wider Render Context

Three observations:

1. Dispersed is still the AI subnet, not a separate network. Per HOGE Wire Sep 3: "Jobs are paid in RENDER, and research shop Messari notes that roughly 95 percent of Dispersed payments are burned, with early clients including Jember, Scrypted and Intelligent Internet." The Dispersed burn fraction (95%) is 3.4x the Sep 16 whole-network burn fraction (27.97%); this means Dispersed is dragging the whole-network burn rate up over time as Dispersed volume grows relative to legacy rendering.
2. The Octane 2027.1 alpha (July 2026) is the demand-side anchor. Per CG Channel: "the headline feature is a neural rendering mode built on NVIDIA's DLSS technology and OTOY's own research." The renderer is now AI-native; every heavy job can be offloaded to Render Network for RENDER credits. This is the structural reason Render's token thesis has converged: the software that generates demand for GPU compute is the renderer itself, and OTOY owns the renderer.
3. The RENDER-as-payment story for AI agents is the open question. Per the June Foundation Report: "millions, tens of millions, billions of these agents interacting online" — and RENDER-based payments are built to fill that gap. The Sep 22 reading is that the agent-payment narrative is still in the "shipping" phase, not yet at measurable volume.

io.net — Steady-State at ~1,245 GPUs, Next Unlock Oct 11

The Aug 25 weekly measured io.net at 71.45B tokens/day on the B200 index (402). The Sep 15 weekly measured the Sep 11 unlock at 14.14M IO ($1.87M, 47.69% supply unlocked). The Sep 22 reading is the Sep 16 12:00/18:00 epochs (1,245-1,251 GPUs, mostly RTX 4090 616 + H100 80GB HBM3 592) and the Oct 11 unlock schedule.

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

1. GPU count is steady-state, not growing. The Sep 16-17 epochs (1,245 / 1,251 / 1,246) are within 0.5% of each other — the network has settled into a 1,200-1,300 GPU range with no clear upward or downward trend. This is consistent with the Aug 25 weekly's note that the io.net GPU-aggregation model has hit its plateau: the network is large enough to attract enterprise customers, but not growing fast enough to absorb new supply without margin pressure.
2. The GPU mix is still 50% consumer silicon. RTX 4090 is 49.4% of the Sep 16 mix (616/1,245); H100 80GB HBM3 is 47.5% (592/1,245). This is the opposite of Gonka's mix (where H100 80GB HBM3 is the median) and reflects io.net's "aggregation" model — anyone with a gaming GPU can run io.net worker software and get paid in IO.
3. The unlock schedule is now monthly at ~$1.20M and trending downward. The Sep 11 unlock was $1.87M; the Oct 11 unlock is $1.20M; the Nov 11 unlock is $1.20M; the Dec 11 unlock is $1.19M. The 2027-07-11 unlock is $1.17M. The step-down to ~$0.53M happens in 2027-08 (3.79M IO); this is the post-cliff supply reduction. The Sep 22 reading is that the unlock overhang is structurally declining month-over-month, which is the most bullish tokenomics signal for io.net through Q1 2027.
4. The IDE burn target (12M IO/year) is ~5% of the monthly unlock schedule. Per globenewswire Jun 11: "io.net to burn up to 12m tokens as network closes 8m deal and hits 4 billion daily AI tokens." At ~1M IO/month burned against ~8.5M IO/month unlocked, the IDE burn removes ~12% of the monthly unlock — which is a meaningful but not deflationary offset. The Sep 22 reading is that the burn rate is sustainable at the 12M IO/year target without requiring additional revenue.
5. The hourly Proof-of-Work (PoW) verification process remains the supply-side integrity layer. Per io.net docs: "Our hourly PoW process runs in the background, causing minimal disruption to normal operations while continuously ensuring the integrity of our network. Suppliers can expect some load on their devices normally for no more than 15 mins per hour to complete the hourly Proof-of-Work authentication process." This is the structural reason io.net can credibly claim ~99% uptime despite being a permissionless GPU aggregator.

Aethir — ATH Circulating Supply Sep 16 Read + 12-Month Roadmap Status

The Sep 15 weekly documented the Sep 9 12-Month Strategic Roadmap. The Sep 22 reading is the ATH circulating supply trajectory through Sep 16 (23.31B ATH, per docs.aethir.com) and the tokenomist.ai burn-feed status (Burn Data Not Available as of Sep 16 20:47 UTC — but the ATH token is on a 42B total supply schedule).

What the table tells you about Aethir's tokenomics:

1. ATH circulating supply is growing at ~0.77B/month (~9.2B/year). The Aug → Sep delta is 766M ATH; the Jul → Aug delta is 766M; the Jun → Jul delta is 766M. The monthly emission is constant at ~766M ATH, which means Aethir's circulating supply will reach 100% of the 42B cap in approximately 24 months (Sep 2028) at the current rate — unless the IDC v2 burn mechanism accelerates meaningfully.
2. The IDC v2 burn + variable platform fee is the offset to the 9.2B/year emission. Per the Sep 15 weekly's coverage of the Sep 9 12-Month Roadmap: "Aethir v2 Mainnet + IDC v2 upgrade in Q1 2026, … burn mechanism + variable platform fee" is now in code terms. The Sep 22 reading is that the burn data is not yet publicly available on tokenomist.ai ("Burn Data Not Available") — which means the IDC v2 burn ledger is either not yet exposing the data or is not yet being scraped. The roadmap calls for "Quarterly Compute Impact Report showcasing enterprise workload metrics for Q1 2026" — which is the next scheduled transparency release.
3. The Aethir Q3 2025 base case (from the Aug 25 weekly) was $39.8M revenue + $166M ARR + ~6.1% Rev/MC ratio. The Sep 22 reading does not yet have the Q3 2026 number — the roadmap calls for "Revenue reports, along with a Quarterly Compute Impact Report" through the end of 2026. The Q3 2026 transparency release is the next inflection point.
4. The Sep 16 tokenomist.ai burn-feed update is a structural data-quality issue. Aethir's burn is happening on-chain but is not yet integrated into the third-party burn tracker — this is the same pattern that Render hit in 2025 (the @RenderBurn tracker is community-built, not first-party). The Sep 22 reading is that Aethir's burn is structurally real but the burn-tracker data is incomplete; the official Q3 2026 transparency report is the authoritative source.

Phala — dstack 0.5.4 Security Anchor + NEUS Trust Layer Production Reference

The Sep 15 weekly documented Phala's 86.5B tokens/day peak (Sep 8) and the 78.09B tokens/day Sep 10 X post. The Sep 22 reading is the dstack 0.5.4 security anchor (the canonical security advisory for the Phala stack as of Q3 2026) and the NEUS x Phala Trust Harness case study (a production reference for using Phala's confidential-compute substrate as the foundation for a verifiable evidence layer).

dstack 0.5.4 (released Sep 1, 2025)

What the dstack 0.5.4 release tells you about Phala's engineering posture:

1. The LUKS header validation CVE is the security anchor. Per the GitHub release: "Fixed LUKS header validation vulnerability (GHSA-jxq2-hpw3-m5wf) — all users should upgrade immediately." This is the second canonical security advisory for the dstack stack (the first being the zkSecurity audit announced Aug 13, per the Sep 8 weekly). The Sep 22 reading is that dstack has two independent security anchors in production: zkSecurity audit + LUKS CVE fix.
2. The 12 new contributors + 70+ PRs is the contributor growth signal. Per the release notes: "12 new contributors joined the project — 70+ pull requests merged." dstack is now a multi-maintainer project, not a single-team codebase. This is the structural reason dstack has crossed from "Phala product" to "TEE runtime" — it has its own contributor base independent of Phala's commercial team.
3. The sdk/rust split into client + types is the no_std enabler. "Separated client and types for no_std compatibility." This is the prerequisite for embedded + WASM dstack clients — the same architectural pattern that lets dstack run inside resource-constrained environments (e.g., TEE-enabled microcontrollers).
4. The verifier release v0.5.11 (current) and the SDK package versioning (dstack-sdk 0.5.4, @phala/dstack-sdk@0.5.8, dstack-sdk 0.5.4b1) show the parallel-tracks architecture. Per the Dstack-TEE/dstack GitHub releases page, the SDK packages version independently of the runtime — which means a user can pin to a stable SDK release while the runtime ships more frequently.

NEUS x Phala: Trust Harness as the Production Reference

The Sep 22 reading includes the NEUS x Phala case study (phala.com/posts/neusphalabuildagenttrustlayer) — a production deployment of an evidence layer on top of Phala's TEE substrate. The architecture has three layers:

1. Sealed inference (Phala layer). "Model calls execute inside Trusted Execution Environments: NVIDIA H100 and H200 GPU enclaves for model execution, Intel TDX confidential VMs for orchestration and pre/post-processing. Prompts, keys, tools, and agent memory are encrypted in memory and decrypted only inside the enclave's protected boundary. The host OS, the cloud provider, and the model operator cannot read the plaintext."
2. Attested runtime (Phala layer). "Before data enters the enclave, Phala's attestation system generates a hardware-backed proof that the runtime environment is genuine, unmodified, and running the expected code. The attestation document binds the enclave's measurement to a specific hardware platform and software hash. Any third party can verify, without trusting Phala or the cloud provider, that inference ran in the exact environment claimed, on the exact code expected, at the time the request was processed."
3. Governed action (NEUS layer). "NEUS sits on top of the inference pipeline and governs who can call it, under what authority, and within what scope. When a team deploys an agent, NEUS issues a trust receipt: a cryptographically signed record that binds the agent's identity to a delegated authority, a scope of permitted actions, and a set of constraints."

The combined artifact: "Two halves of one proof: the data was never exposed, and the action was authorized." This is the production case study that confirms the Phala pitch — confidential inference + on-chain receipts + third-party verification — works end-to-end with real customers (NEUS, OpenRouter, Z.AI, OPPO, Venice AI, NEAR AI, Uniswap per the case study).

Phala Production Customer List (per NEUS case study)

Akash — Razer AVA Mini Case Study + Sep 15-17 AI Infra Summit Recap

The Sep 15 weekly noted the AI Infra Summit opens TODAY (Sept 15-17, Santa Clara) and that DePIN-native networks are not on the main stage. The Sep 22 reading is the Razer AVA Mini case study (the production reference Akash published alongside the AI Infra Summit) and the post-summit coverage.

Razer x Akash — AVA Mini Production Case Study

Per the akash.network homepage (Sep 22 read):

The Razer case study is structurally important because it is the first consumer-scale product reference for Akash: Razer is shipping the AVA Mini campaign to end-users, and the inference is running on Akash. The $0.01/image cost is the cost anchor for consumer-AI workloads — it's the same price point as Midjourney's lowest tier, and 15x cheaper than running the same workload on a centralized API.

AI Infra Summit Post-Summit Read

The Sep 15-17 Santa Clara summit was sponsored by NVIDIA, Intel, Meta, OpenAI, Google, Dell, and HCLTech (per the Sep 15 weekly). The Sep 22 reading is the post-summit takeaway: DePIN-native compute networks remained on the expo floor and breakout sessions, not the main stage. The Crusoe agentic-inference breakout and the Guild AI breakout are the only DePIN-relevant sessions. The Sep 22 implication is that the "DePIN is invisible to enterprise buyers" narrative from the Aug 25 weekly (Stripe buys OpenRouter for $7.5B) continues to hold — DePIN inference is winning on cost (15x cheaper per the Razer case study) but not yet winning on enterprise mindshare.

Akash Q4 2026 Roadmap Status (Sep 22 read)

The Sep 22 reading is that AEP-85 Buildpacks-Powered Deployments is the next near-term commitment (Sep 30). Per the AEP/ROADMAP.md: this is the developer-experience improvement that closes the gap with Phala's phala deploy CLI.

Fluence — VPC Live Sep 8, FLT Floor $2.09M Sep 6

The Sep 8 weekly flagged the Fluence Spheron GPU + FLT-stables thread. The Sep 22 reading is the Fluence VPC release (Sep 8) and the post-release pricing context.

Fluence VPC (Sep 8, 2026)

The VPC release is structurally important because it is the first "non-GPU" infrastructure release of 2026 for Fluence — every prior 2026 release was GPU-focused (Spheron partnership, FLT-stables thread, GPU auctions Aug 4). VPC is the missing enterprise primitive that lets Fluence host multi-tier applications (database + API + worker on separate VMs in a private subnet) instead of just GPU containers.

Fluence Network Stats (Sep 22 read)

The Sep 22 reading is that Fluence's narrative is now GPU + VPC, not just GPU. The Sep 6 FLT market cap of $2.09M is the price floor the network has been trading at since the Aug 25 weekly; the VPC release is the structural argument that FLT should re-rate, but the price hasn't yet moved.

Golem — GLM Breaks Resistance Sep 13 (857K Volume, +17.36% Week)

The Sep 15 weekly noted the GLM migration invariant holds (796,565,516.94 GLM + 203,434,483.06 GNT = 1B). The Sep 22 reading is the Sep 13 resistance breakout:

The breakout is structurally interesting because the 7-day price change (+17.36%) is roughly 2x the volume surge (2x spike vs 7-day average) — meaning the price move is driven by genuine demand, not by a single whale. The Sep 22 reading is that GLM has cleared the 0.1148 resistance that held for the prior 30 days and is now testing the 0.1263 short-term resistance. If 0.1323 breaks, the next leg toward the prior Q2 2026 peak of ~$0.16 opens.

The structural backdrop for GLM:

1. Golem-JS SDK 2.0 (released 2025) is the developer anchor. Per CMC AI Sep 9: "Golem-JS 2.0 SDK Release (2025) – A foundational upgrade adding invoice management, structured error handling, and critical stability fixes."
2. NetMind AI deploys GLM-5.3 Flash at $0.075/M input. Per CMC AI Sep 9: "GLM 5.3 Flash is already live on NetMind API. $0.075 / 1M input... 1M context. Native multimodal. 600 RPM." — NetMind is one of the first major API providers to deploy Z.ai's flagship reasoning model.
3. Salad.com enterprise trial (Q1 2026) continues. Per CMC AI Sep 9: "Salad Compute Pilot Scaling (Q1 2026) – Ongoing enterprise trial to validate Golem's network for large-scale commercial GPU workloads."
4. GLM/BTC delisted from Binance (Jul 14). Per CMC AI Sep 9: "Binance Delists GLM/BTC Pair (14 July 2026) – The exchange removed the trading pair due to low liquidity, a near-term headwind." GLM/USDT remains active.

The Sep 22 reading is that the GLM breakout is the first price-action confirmation that the developer-anchor (Golem-JS 2.0) + enterprise-trial (Salad) + AI-model-deployment (NetMind + GLM-5.3 Flash) narrative is starting to be valued by the market.

Other Notes (Short)

• Marlin Oyster — steady-state; EC2 attestation SDK + x402-gateway remain the active platform surfaces. No new items this week.
• Bacalhau — repo push cadence steady. NCL Protocol remains default per Sep 1 weekly. The compute-over-data architecture documentation is now live at bacalhau.org/docs (Sep 22 read confirms the "Keep the network boundary explicit" guidance: object storage is not automatically mounted just because a node is nearby, and a job constraint does not change a storage bucket policy).
• Prime Intellect — no public release this week. The Jul 8 Series A close ($130M, $1B valuation, Radical Ventures lead) remains the freshest data point; the Sep 7 push is the most recent code activity.
• Bittensor Exploit Summit — Sep 28-29 confirmed (Sep 15 weekly). The Sep 22 reading adds V461 Basket Trading testnet release as the new spec the room will be arguing about.
• Gensyn — RL Swarm paused per Sep 1 weekly. No new updates this week.
• Decentralized training — INTELLECT-2 globally distributed RL remains the active research reference (Sep 8 weekly). The Jul 8 Prime Intellect Series A close remains the freshest commercial reference.

Evidence at a Glance

What to Watch This Week

Primary Sources

• 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
• Gonka network updates Sep 11-16 — gonka.ai/docs/network-updates
• Gonka live network state Sep 19 — gonka.gg/network
• Gonka Proposal 99 (TOKEN2049) — gonka.ai/docs/network-updates#september-4-2026
• Gonka hardware specifications — gonka.ai/docs/host/hardware-specifications
• Render burn tracker Sep 16 — @RenderBurn tweet captured by CMC AI
• Render Network August 2026 Report — thechainobserver.com/render-network-publishes-august-2026-report
• Render Network 2026 strategic shift (HOGE Wire Sep 3) — hoge.gg/render-2026-gpu-network-building-its-own-demand
• 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
• io.net vesting schedule — coinlaunch.space/projects/ionet/vesting
• io.net PoW verification process — io.net/docs/guides/workers/proof-of-work
• 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
• dstack 0.5.4 release — github.com/Dstack-TEE/dstack/releases/tag/v0.5.4
• dstack releases — github.com/Dstack-TEE/dstack/releases
• 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 Trust Center — trust.phala.com
• Akash Razer AVA Mini case study — akash.network
• Akash 2026 Roadmap — akash.network/roadmap/2026
• 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
• 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