Decentralized Compute Tuesday — September 15 2026 Weekly Update

Weekly covering September 9 – 15, 2026. Five structural threads this week. First: Phala's daily-token figure roughly doubled in eight days — the home-page server-render climbed from 36.7B (Sept 13 read) to 86.5B on Sept 8 (one week earlier per the server-render timestamp visible on phala.com) and 78.09B on Sept 10 (per the PhalaNetwork X post) — with the Sept 10 mix being GLM-5.3 at 43.54% + GLM-5.3 Flash at 26.96% = Z.ai's stack carrying 70.5% of all Phala throughput. Second: Aethir published its 12-Month Strategic Roadmap on Sep 9 — making concrete what "burn mechanism + variable platform fee" means in code: Aethir v2 Mainnet + IDC v2 upgrade in Q1 2026, EigenLayer ATH Vault, an explicit chain migration planned for 2026, Aethir Developer SDK + Compute Reputation Layer + AI Workload Marketplace v2 + zk-proof compute-validation prototype in Q2 2026, Compute-as-a-Service (CaaS) pricing in H2 2026, and Aethir v3 + AI orchestration APIs late 2026. Third: the AI Infra Summit opens TODAY (Sept 15–17, Santa Clara) — the speaker list confirms DePIN-native compute networks are not on the main stage, with Intel (Lip-Bu Tan), NVIDIA (Ian Buck), Meta (Santosh Janardhan), OpenAI (Sachin Katti), Google (Dave Patterson), Dell (John Roese), HCLTech (C Vijayakumar) commanding the keynote program; the only DePIN-relevant sessions are expo-theater breakout talks (Crusoe on agentic inference; Guild AI; etc.). Fourth: Exploit Summit confirmed speaker lineup (Sep 9 X update) adds Jacob Steeves (Affine CEO) as a new entrant — Affine is the full-stack confidential-compute angle on Bittensor, joining Jon Durbin (Chutes/SN64), Rob Myers (Manifold), Marcus Graichen (Taostats). Fifth: Akash's roadmap commits get to date — AEP-67 Console Bid Screening moved from "Last Call since May 2025" to delivered Jul 17 per the on-chain AEP/ROADMAP.md, and AEP-46 Spot Instances is delivered Aug 30; the only remaining Q4 commitments are AEP-36 (Custom Domain Configuration Nov 15), AEP-11 (Managed Services Market Dec 15), AEP-54 (NextGen AMD GPU Dec 15), and the AEP-79 (Shared Security Dec 31) decision point that closes the Cosmos-to-Solana migration debate. Below: io.net's Sep 11 14.14M IO unlock (Community tranche, 47.69% supply now unlocked), Prime Intellect's $130M Series A + $1B valuation context, Marlin Oyster steady-state, Bacalhau + Fluence quiet.

Macro Thread — Phala's 78B Token Day: Z.ai's Stack Now Carries 70.5% of All Phala Throughput

The most important structural observation from this week's Phala reads: the supplier mix has narrowed onto Z.ai's stack in a way that no other DePIN inference network has matched, and the underlying token totals are now the largest measured on any DePIN inference surface.

What the four-row table tells you:

1. Phala cleared 78B tokens/day twice in seven days. The Sep 8 read of 86.5B (visible as the home-page 2026-09-08 timestamp) is the largest single-day server-render reading on record for any DePIN inference network. It is +113% above the Sep 2 reading of 40.4B flagged in the Sep 8 weekly, and +136% above the Sep 13 reading of 36.7B.
2. The mix tells you why. GLM-5.3 (Z.ai flagship reasoning model) alone accounts for 43.54% of all Phala throughput on the Sep 10 day; GLM-5.3 Flash (Z.ai multimodal 320B MoE with 18B active params, Aug 27 launch on Phala) carries another 26.96%; together Z.ai's two-SKU stack represents 70.5% of all Phala daily tokens. The remaining 30% is split across DeepSeek V4 Flash, Kimi K3, Qwen3.6/3.8-27B, Muse Glimmer 30B, Gemma 4 31B, and others. This is the single tightest supplier concentration in DePIN inference and is structurally different from AkashML's flatter spread (which the Sep 8 weekly measured at 23.6B/week #6 on OpenRouter) or Aethir Mesh LLM (which is mostly B200/H200 token throughput, not OpenRouter's provider page).
3. Why Z.ai's stack resonates with Phala specifically. GLM-5.3 (released Aug 14, 743B base model, post-training gains) is a long-horizon coding/agent model with a 1M-token context window; GLM-5.3 Flash is the multimodal 320B MoE variant with 18B active params. Phala's confidential-AI pitch is the privacy-preserving route for enterprise code agent and document analysis workloads — which is exactly the Z.ai value proposition. Z.ai's stated reasoning: developers want open-source SOTA capability + OpenAI-compatible routing + sealed private inference. Phala is the only DePIN network hitting all three.
4. The pattern is also visible on Phala's models page. Sep 13 reads 12 distinct SKUs across Z.ai, Qwen, Meta, DeepSeek, MoonshotAI, NVIDIA, Google (Nemotron 3.5 Lightning at $0.08/M input is the price-floor SKU; Kimi K3 at $3.00/M input is the ceiling). The catalog is the deepest open-weight DePIN menu on OpenRouter by SKU count; AkashML is at 6.
5. What is the steady-state? The four readings (40.4B, 86.5B, 78.09B, 36.7B) tell you Phala's home page captures a most-recent rolling peak and that intra-week variability is large. The conservative floor is now ~36-40B/day (vs ~28B in the Sep 1 weekly's reading); the rolling-peak ceiling is 86.5B. The Phala team's Sep 10 X post is the first time they have publicly claimed a specific 24h-token figure tied to a model-mix breakdown.

What the Phala Trajectory Means for the DePIN Inference Stack

Three observations:

1. Phala is the only DePIN inference network with confirmed 70B+ daily-token days. AkashML tops out at ~24B/week (#6 on OpenRouter per Sep 1 weekly). Aethir Mesh LLM and io.net IO Intelligence route through separate surfaces. Gonka's measured daily volume (per crypto.news coverage) is much lower. Phala's Stripe/OpenRouter deal — closed Jul 18 per the Aug 4 weekly — is the gating event that explains the 28B → 86B ramp in 60 days.
2. Z.ai as the upstream lab is a structural choice, not a market accident. Z.ai delayed the GLM-5.3 weights two weeks for safety testing (Axios Aug 14: "China's Z.ai holds GLM 5.3 release over hacking risks" — 84.5% CyberGym, 2,400 security flaws found to date). The model is the strongest open-weights reasoning model on public coding benchmarks (Terminal Bench 3.0: 4.6 → 28.3 vs GLM-5.2; DeepSWE 46.2 → 66.9; Agents' Last Exam CLI 23.8 → 28.5). The pitch is "best open weights + best open-source SOTA + best confidential inference" — and Phala owns the third leg exclusively.
3. The audit overhang is now fully cleared. Phala's zkSecurity audit (announced Aug 13, first OS-level TEE platform audit) plus SOC 2 Type I, HIPAA, and ISO 27001-in-progress are the buy-side checklist. The Sep 8 weekly noted the structural critique ("audit was the open question") is now answered.

Aethir — 12-Month Strategic Roadmap Operationalizes the "Burn + Variable Fee" Commitment

The Sep 8 weekly noted "IDC tokenomics change is now in code terms, not just narrative." The Sep 9 Aethir blog post turns that into a dated roadmap with concrete deliverables. Reading the Q1 2026 section against the Sep 8 "burn + variable fee" line item is the single most clarifying document Aethir has published since the ACCELERATE launch.

What the roadmap answers structurally:

1. The "burn + variable fee" line item from the Sep 8 weekly now has four concrete operational forms. (a) Q1 2026's IDC v2 introduces "full on-chain transparency for service fees and rewards" — so the variable fee becomes a transparent on-chain parameter rather than an opaque Foundation discount. (b) Q2 2026's AI Workload Marketplace v2 + zk-proof validation prototype means the burn can be gated by zk-verified compute epochs (the natural counterpart of the Q2 zk-proof R&D). (c) H2 2026's CaaS pricing is the recurring-revenue surface where the burn acts as a deflating counterweight to enterprise ARR — making ATH's relationship to ARR explicit. (d) Aethir v3 late 2026 ships the multi-chain compute dashboard so ATH tracks utilization across deployments.
2. The chain migration is the unannounced structural risk. The roadmap lists it as a 2026 commitment without naming the destination chain, the migration mechanics, or the holder-migration terms. Akash's analogous AEP-79 Shared Security decision (Dec 31 target) is the only parallel in DePIN compute right now. Two simultaneous 2026 chain decisions in two DePIN compute networks is a sector-defining moment.
3. The Sep 7-14 blog cadence is unusually high — six pieces in seven days. Sep 7 Neocloud Debt and the GPU Collateral Value Problem; Sep 8 ACCELERATE Uses Demand-Led Capacity Planning; Sep 9 The Anthropic-Lambda Deal and the AI Compute Market; Sep 9 Aethir's 12-Month Strategic Roadmap (above); Sep 14 What NVIDIA's Hugging Face Acquisition Changes. This is the most aggressive Aethir content cadence in 2026 and reads as the structural answer to "ATH's value accrual low for no fee distribution and no buy-and-burn" critique that has defined the ATH bear case since the Coinbase Sep 3 perp delist.

Counter-Reading — CaaS Pricing is the Most Important Single Line in the Roadmap

The H2 2026 Compute-as-a-Service pricing model entry is structurally the most important line on the roadmap because:

1. Recurring-revenue contracts paid through the Strategic Compute Reserve make ATH an institutional-grade asset class for the first time. Today, the SCR holds $344M of staked ATH; with CaaS in place, that stake backs enterprise GPU bookings on a multi-quarter cadence — the kind of demand floor that turns ATH from a DePIN utility token into a Treasury-style reserve asset.
2. Hybrid cloud partnerships ("Oracle, Tencent, AWS") are the most aggressive distribution statement Aethir has ever made. The Sep 8 weekly noted Aethir was publishing positioning copy against Stripe/OpenRouter. The roadmap's "partner with major clouds" line is the explicit answer: Aethir is positioning to sell to enterprise as a GPU commodity (through Oracle, Tencent, AWS) rather than only to crypto-native clients. The implication for Akash, io.net, and Render is that the competitive surface for enterprise compute now includes legacy cloud channels.
3. Aethir v3 + AI orchestration APIs in late 2026 ship the multi-chain compute dashboard. That is the moment Aethir becomes directly comparable to AkashML (which has been selling through OpenRouter for 8 months at #6). Whether Aethir ends up inside OpenRouter's catalog or only sells direct is a Q3-Q4 2026 decision to watch.

Bittensor — Exploit Summit Speaker Lineup Adds Jacob Steeves (Affine); AI Infra Summit Opens Today Without DePIN on Main Stage

Two events bracket this week's Bittensor-relevant industry activity:

What the speaker lists tell you:

1. AI Infra Summit confirms DePIN is absent from the hyperscaler-track conversation. The main stage is dominated by NVIDIA + Intel + AMD + hyperscaler system integrators + Meta + OpenAI + Google + Dell + HPE + HCLTech. The closest DePIN-adjacent names are Crusoe (a neocloud that has been positioning against DePIN by emphasizing its vertically-integrated data center footprint) and Marlin (which sponsors breakouts but is not on the main stage agenda). The takeaway for decentralized compute positioning is: AI infrastructure is being framed as a chip + system integrator + hyperscaler + neocloud conversation, not a DePIN one. Bittensor / Phala / Akash / Render have to win the Sep 28-29 Exploit Summit conversation as the only counter-stage to that narrative.
2. Jacob Steeves (Affine CEO) joining Exploit Summit is the structural Affine moment. Affine is full-stack confidential compute on Bittensor — Steeves's framing per the exploitsummit.com speaker page is "decentralized confidential compute into production-grade systems enterprises can trust." Rob Myers / Manifold runs SN4 (Targon); Affine runs adjacent confidential-compute subnets. The Sep 10 weekly (WoT) noted that Bittensor's institutional quiet through August was the dominant signal; the Sep 28-29 lineup with Steeves is the first time confidential-compute builders are formally joined under one stage. The Yuma-consensus + subnet-economic structure of Bittensor is now structurally adjacent to Phala's dstack architecture — the cross-pollination window is the Sep 28-29 conference.
3. Marcus Graichen (Taostats founder) joins. Taostats is the de-facto Bittensor ecosystem analytics surface. With Graichen on stage, the institutional-data conversation (subnet emission, validator scoring, dTAO distribution) gets a public-facing channel that has been missing from the Bittensor press surface for most of 2026.

GTAO Window Holds; Bitwise vs Grayscale S-1 Amendment Status

The Sep 8 weekly noted the Grayscale GTAO S-1 window was open with no SEC ruling. The Sep 10 confirmation (youtube video "Grayscale Just Filed for a Bittensor ETF — TAO Is Next"): Grayscale filed Amendment No. 1 to their S-1 on Apr 2, 2026 moving the GTAO Trust toward a spot ETF on NYSE Arca. The Bitwise TAO Strategy ETF (60% TAO + rest TAO ETP) is in flight. No SEC procedural ruling this week; both filings remain in window. TAO market cap is now ~$3.36B per the Sept 10 YouTube coverage.

Akash — Roadmap Corrections: AEP-67 Delivered Jul 17 (Not "Last Call Since May 2025"); AEP-46 Delivered Aug 30

The Sep 8 weekly listed AEP-67 Console Bid Screening as "Last Call since May 2025" and AEP-46 Spot Instances as "Draft since Dec 2024." The canonical AEP/ROADMAP.md at akash-network/AEP/blob/main/ROADMAP.md shows the published completion dates are firm:

Three corrections from the Sep 8 weekly:

1. AEP-67 has been delivered since Jul 17 — not "Last Call since May 2025." The Sep 8 weekly's claim was based on the akash.network/roadmap/aep-67 page formatting that still says "Last Call"; the canonical AEP/ROADMAP.md at the repo gives 2026-07-17 as the completion date. This is consistent with the Sep 1 weekly's confirmation that AEP-67 was delivered alongside AEP-46 on Aug 30/31; the dates in the repo are the source of truth.
2. AEP-46 is delivered (Aug 30) — not "Draft since Dec 2024." The AEP's spot-instance design is now in production.
3. AEP-85 (Buildpacks-Powered Deployments) is a new entry for Sep 30 that was not in the Sep 8 weekly. Buildpacks is the Heroku-style automatic buildpack orchestration that lets developers push source code without writing Dockerfiles; combined with AEP-36 (Custom Domain Nov 15) and AEP-11 (Managed Services Market Dec 15), the Q3-Q4 2026 Akash roadmap is now a three-pillar developer-experience buildout: buildpacks (Sep 30), custom domains (Nov 15), managed databases (Dec 15).

What this means for Akash's competitive position:

1. The Sep 30 AEP-85 deadline is the next concrete checkpoint. Buildpacks is the gateway from "developer with Dockerfile" to "developer with git repo" — the lowest-friction developer path in DePIN compute. Phala already ships phala deploy -c docker-compose.yml; Akash matching that with Buildpacks is the structural race for "easiest DePIN inference platform to deploy to."
2. AEP-79 Shared Security Dec 31 is the Cosmos-to-Solana decision point. The AEP is the formal home for the migration debate; if it ships by Dec 31 with a "stay on Cosmos" stance, that's a binding network decision. If it ships with a Solana migration plan, Aethir's parallel chain-migration becomes the second high-profile 2026 migration in DePIN compute.
3. The competitive surface against Phala. Phala leads on confidential AI inference (10+ models on OpenRouter, 86.5B/day peak); Akash leads on general-purpose Kubernetes workloads and is about to ship the developer-experience gap closer (AEP-85 Buildpacks). The two networks are complementary rather than competitive on most workloads, but the enterprise developer audience that wants private LLM inference + standard k8s manifests + domain-named endpoints + managed databases has a real path through both.

io.net — Sep 11 14.14M IO Unlock (Community Tranche, 47.69% Supply Now Unlocked)

Three observations about io.net's standing this week:

1. 14.14M IO unlocked Sep 11 (Binance Square Sep 4 confirmation; Tokenomist tracking; CMC AI Sep 11 entry). The unlock was valued at ~$1.38M (~2.59% of mcap at the time), released to the Community tranche. 47.69% of total IO supply is now unlocked (381,483,051 IO out of ~800M total). Weekly unlocks continue through the schedule; the Sep 8-13 cohort included 14 altcoins (Tokenomist / CMC AI).
2. IDE burn target holds. The 12M IO target burn for year-one (announced Jun 11) remains the dominant supply-side narrative. CoinDesk's research note "The Incentive Dynamic Engine: A New Era for io.net Tokenomics" frames IDE as "ties emissions and burns directly to network earnings" — making the burn revenue-tied, not fixed-rate. io.net's Aug 13 blog "The GPU crisis won't be solved by building more data centers" and Aug 18 "What happens when you can't get a GPU? The hidden cost of cloud wait times" are the public stance.
3. Mint-immutability confirmed Sep 8 (Own-Token independent RPC read, carried forward from Sep 1 weekly). Supply at 798,538,520.02 IO; mintAuthority = NULL; freezeAuthority = NULL.

Prime Intellect — Series A Closed Jul 8; $130M Raise + $1B Valuation + Radical Ventures Lead

The Prime Intellect research cadence (Sep 7 repo push; Sep 6 ARC-AGI-3 press) was covered in the Sep 8 weekly. New structural context from RL-List.com vendor profile:

What the funding detail tells you:

1. Prime Intellect is the highest-valued decentralized-training network as of mid-2026. The $1B Series A valuation (Jul 8) places it above Prime Intellect's earlier peer Gensyn (which priced its token in 2025 at a much lower valuation per tokenomist estimates) and Bittensor's market cap of $3.36B (Sep 10 YouTube coverage — but Bittensor is a network, not a Series A company). The Jul 8 close means Prime Intellect has 18-month operational stability through the next research cadence: INTELLECT-4 or Prime Agent v2 or the next RL environment push.
2. The angel list is the deepest researcher-aligned roster in DePIN compute. Karpathy / Schulman / Tri Dao / Dylan Patel / Delangue is the strongest possible researcher signal. Combined with Radical Ventures (the AI-focused firm behind Cohere and other frontier-AI bets), the structural backing positions Prime Intellect as the "researcher-funded" answer to centralized frontier labs — and the Sep 8 weekly's note about Zapier + Ramp customer stories is the commercial counterpart.
3. No public release cadence in this weekly window. The Sep 7 push is the most recent; the Sep 6 ARC-AGI-3 coverage was the prior anchor. Prime Intellect's weekly cadence is research-output-led, not press-output-led. The next public research item is likely late-Sep or early-Oct.

Other Notes (Short)

• Marlin Oyster — steady-state; EC2 attestation SDK + x402-gateway are the active platform surfaces from the Sep 8 weekly. No new items this week.
• Bacalhau — repo push cadence steady. NCL Protocol remains default per Sep 1 weekly.
• Fluence — FLT market cap Sep 6 read ($2.09M, #2284) holds as the price floor. The Spheron GPU + FLT-stables thread from the Sep 1 weekly is still the narrative bright spot but not yet price-visible.
• Golem — Aug 13 GLM migration invariant holds (796,565,516.94 GLM + 203,434,483.06 GNT = 1B). The Salad GPU engineering trial remains the only narrative-active thread.
• Gonka — crypto.news Sep 11-12 coverage continues the useful-PoW / inference-network positioning. Dahl live inference tracker (inference.dahl.global/gonka-network) is the public measurement surface.

Evidence at a Glance

What to Watch This Week

Primary Sources

• Phala home-page server-render (Sep 8 = 86.5B, Sep 13 = 36.7B) — phala.network
• PhalaNetwork X post Sep 10 — 78.09B tokens, GLM-5.3 43.54% + GLM-5.3 Flash 26.96% — x.com/PhalaNetwork/status/2098088689067008476
• Phala GLM-5.3 Flash model page — phala.com/models/z-ai/glm-5.3-flash
• Z.ai GLM-5.3 launch (Axios Aug 14) — axios.com/2026/08/14/china-open-source-ai-glm-53
• Z.ai GLM-5.3 blog post — z.ai/blog/glm-5.3
• Aethir 12-Month Strategic Roadmap (Sep 9) — aethir.com/blog-posts/aethirs-12-month-strategic-roadmap-supercharging-enterprise-ai-compute-growth
• Aethir Anthropic-Lambda Deal (Sep 9) — aethir.com/blog-posts/the-anthropic-lambda-deal-and-the-ai-compute-market
• Aethir Neocloud Debt (Sep 7) — aethir.com/blog-posts/neocloud-debt-and-the-gpu-collateral-value-problem
• Aethir ACCELERATE Demand-Led Capacity (Sep 8) — aethir.com/blog-posts/aethir-accelerate-uses-demand-led-capacity-planning
• Aethir NVIDIA Hugging Face (Sep 14) — aethir.com/blog-posts/what-nvidias-hugging-face-acquisition-changes
• AI Infra Summit 2026 agenda — ai-infra-summit.com/2026agenda
• AI Infra Summit 2026 about — ai-infra-summit.com/about
• NVIDIA AI Infra Summit event page — nvidia.com/en-us/events/ai-infra-summit
• Exploit Summit 2026 speakers — exploitsummit.com/speakers
• Bittensor Commons Exploit Summit confirmation (Yahoo FR) — fr.finance.yahoo.com/actualites/bittensor-commons-confirme-lévénement-exploit-161900692.html
• Akash AEP/ROADMAP.md — github.com/akash-network/AEP/blob/main/ROADMAP.md
• AEP-67 Console Bid Screening — akash.network/roadmap/aep-67
• AEP-46 Spot Instances — akash.network/roadmap/aep-46
• AEP-36 Custom Domain Configuration — akash.network/roadmap/aep-36
• io.net 14.14M unlock Sep 11 (Binance Square) — binance.com/en/square/post/09-04-2026-io-net-will-unlock-approximately-14-14-million-io-tokens-on-september-11th-362907395996775
• io.net Tokenomist vesting — tokenomist.ai/io
• io.net CMC AI Sep 11 entry — coinmarketcap.com/cmc-ai/io-net/latest-updates
• io.net IDE / 12M IO burn (Jun 11) — globenewswire.com/news-release/2026/06/11/3310232
• io.net IDE research note (CoinDesk) — coindesk.com/research/the-incentive-dynamic-engine-a-new-era-for-io-net-tokenomics
• io.net blog — io.net/blog
• Prime Intellect Series A vendor profile — rl-list.com/vendors/prime-intellect
• Prime Intellect research cadence — primeintellect.ai
• Phala CMC AI entry (Sep 14) — coinmarketcap.com/cmc-ai/phala-network/latest-updates
• Grayscale GTAO S-1 / Bitwise TAO ETF — tradingview.com/news/coinpedia:861fc7055094b
• Gonka Dahl live inference tracker — inference.dahl.global/gonka-network
• Golem GLM migration invariant — ownyourmind.ai/projects/golem
• Fluence FLT Sep 6 read — bybit.com/en/price/fluence-2