Weekly covering September 2 – 8, 2026. Five structural threads this week. First: Phala hit a new peak at 40.4B tokens/day on the Sept 2 home-page snapshot (vs 28.1B at the Sep 5 server-rendered reading) — a +44% intra-week spike with 10 active models, evidence that the Stripe/OpenRouter provider listing continues to drive demand on the most measurement-rich DePIN inference network. Second: Aethir shipped its biggest structural commitment of 2026 — ACCELERATE's ten-site US/EU build (up to 20 MW, B300/GB300, $2B target) was confirmed in an Aug 24 announcement that lands inside this weekly window in its tokenomics impact, and on Sep 7 ME News carried the formal "burn + variable fee" framing. Third: Bittensor's institutional window widened with the Exploit Summit Sept 28–29 Montreal lineup (Jon Durbin / Chutes, Rob Myers / Manifold, 100+ subnet teams) plus the AI Infra Summit Sept 15–17 Santa Clara as a co-incident industry stage; Grayscale GTAO S-1 decision still in window, no SEC ruling. Fourth: Akash's Q4 2026 roadmap publishes four firm dates — Custom Domain Config (Nov 15), Managed Services Market (Dec 15), Onchain Provider Incentives (Dec 31), Shared Security (Dec 31) — alongside AEP-46 Spot Instances still Draft and AEP-67 Console Bid Screening now Last Call. Fifth: dstack silently re-orged — the canonical runtime moved from Phala-Network/dstack-cloud (archived Jun 10) to Dstack-TEE/dstack, with the cloud-wrapper archive vs runtime split now reflected across the documentation surface. Below: io.net IDE burn-rate context, Prime Intellect post-INTELLECT-3 research cadence, Marlin Oyster EC2 attestation SDK surface, Bacalhau Sept 6 push, Fluence FLT/stables and Golem GLM migration invariant.
Macro Thread — Phala Pushes Past 40B Tokens/Day, OpenRouter Provider Ranking Hardens
The week crystallizes a single observation that has been building since the Stripe/OpenRouter deal: the public provider page is the demand surface, and Phala is currently the highest-leverage network on it.
| Date | Phala home-page daily tokens | Active models | Source |
|---|---|---|---|
| Aug 22 | 37.11B (peak before Sep) | ~7 | Aug 25 weekly |
| Sep 2 | 40.4B | 10 | phala.network server-render |
| Sep 5 | 28.1B | 10 | phala.network server-render |
What the table tells you:
- The Sept 2 reading is the largest single-day total Phala has reported on its home-page server-render. Within three days the read drops to 28.1B; the most plausible reading is that Phala's home page captures the most-recent rolling-window peak, and the peak on Sept 2 was 40.4B while the steady-state on Sep 5 was 28.1B. Either way, Phala cleared the 40B-tokens/day ceiling for the first time.
- Model count grew from ~7 (Aug 22) to 10 (Sep 5). The new SKUs that drove the bump were the GLM-5.2 / DeepSeek V4 Flash family plus the Qwen3.7/3.6/3.8-Flash-Next Chinese open-weight wave; the catalog is now deeper than any other DePIN inference network on OpenRouter (vs AkashML's 6, which stayed flat).
- The Phala home-page shows a hard floor: 10 active models, ~28B/day, two-week rolling peak. Aethir-side, AkashML, and Gonka all show zero on this surface — they don't sell through OpenRouter. io.net's IO Intelligence gateway is a separate SKU stack. The fact that only Phala, AkashML, and a handful of centralized providers (Venice, Friendli, Groq, Cohere, Reka) appear in the top quartile means Phala is the only DePIN-native network currently hitting Groq-class last-week token volumes inside a TradFi routing surface.
- The audit overhang is now structurally cleared. zkSecurity's May–June 2026 dstack audit (announced Aug 13 —
Historic moment for confidential computing. dstack just passed its first-ever security audit by @zksecurityXYZ.) was the first OS-level TEE platform audit; by the time SEC and SOC 2 buyers come asking, the answer is on the public record. The Phala team has now mapped SOC 2 Type I, HIPAA, and ISO 27001-in-progress onto that audit baseline.
The dstack Re-org Is Now Visible on GitHub
Three observations about the repo state:
Phala-Network/dstack-cloudwas archived by the owner on Jun 10, 2026 — read-only, 6 stars, 2,263 commits, archived with no successor pointer in the README. Anyone landing on it via search today sees an "archived" banner.- The canonical runtime is
Dstack-TEE/dstack— the GitHub org rename is consistent with thego get github.com/Dstack-TEE/dstack/sdk/goimport path now published on the dstack landing page. The cloud-wrapper layer (dstack-cloud) and the runtime layer (dstack) are now visibly distinct GitHub entities. - The attestation-pipeline hardening article dated Jun 16 (Phala blog, tagged Reports / dstack) ships a fix track that the earlier tracking systems undercounted. Anyone tracking dstack via the archived repo alone would have missed Jun 16's attestation work.
For security auditors and CI consumers: pin to Dstack-TEE/dstack (not Phala-Network/dstack-cloud) when wiring in 2026 H2.
Aethir — ACCELERATE Confirmed, IDC Tokenomics Has Burn + Variable Fee
Aethir made the biggest structural commitment in its history on Aug 24, with details now landing in this weekly window in tokenomics terms:
| ACCELERATE commitment | Value | Source |
|---|---|---|
| Sites | 10 in US + Europe | aethir.com/blog-posts/aethir-launches-accelerate |
| Power | up to 20 MW | same |
| Hardware | NVIDIA B300 + GB300 clusters, 64–256 nodes each | ownyourmind.ai (Aug 24 buildout) |
| 2026 contract target | up to $700M | same |
| Full-build contract target | over $2B | aethir.com |
| Tokenomics change | burn mechanism + variable platform fee | aethir.com + ME News Sep 7 |
| Coinbase ATH perp delist | Sep 3 | ownyourmind.ai |
What changed since the Aug 25 weekly:
- The IDC tokenomics change is now in code terms, not just narrative. Aethir's own Aug 24 announcement says the updated policy "will include" a burn mechanism and a variable platform fee. ME News carried the "burn mechanism + variable platform fee" framing again on Sep 7 as part of the "ACCELERATE Beats 2026 AI Capacity Delays" coverage. The introduction of a variable fee is the structural counterpart to the burn: provider earnings become volume-sensitive, and the Foundation can tune fee level to absorb ATH price volatility without forcing per-epoch payout renegotiation. The Own Your Mind August assessment (
ATH's value accrual low for having no fee distribution and no buy-and-burn, so on paper this is the fix) reads as the structural critique Aethir is responding to. - Coinbase ATH perp delist on Sep 3. Coinbase listing standards reviewed IOTX in the same fortnight (delist Sep 23, named Badger DAO + Storj as previously removed); ATH's review came first. ATH is at ATL on the spot chart; the delist removes a perpetual-liquidity venue that institutions use to hedge. The tokenomics refresh is now the only remaining structural answer.
- The Sep 2 Aethir community article "AI Model Routing in 2026: Inside the Stripe & OpenRouter Deal" is a direct piece of competitive content. Aethir is publishing positioning copy for its $2B buildout against the new reality that inference demand routes through a TradFi-owned aggregator. Whether Aethir ends up inside or outside OpenRouter's catalog is a Sep/Oct decision to watch.
Counter-Critique That Has Not Aged Well
Axe Compute's prepay + 45-day-build model (Q2: $317M prepayments, $3.2B signed, per the Aug 25 weekly) requires ATH at a level it is no longer at. The ownyourmind.ai assessment that Aethir's review scores ATH's value accrual low for "no fee distribution and no buy-and-burn" was made when ACCELERATE was already announced but before the tokenomics refresh was formalized. With burn + variable fee now in the protocol roadmap, the structural critique has a concrete response — the question is whether the response arrives before prepay-funded site builds outpace token-minted reward obligations.
Akash — Q4 2026 Roadmap Publishes Four Firm Dates
The Akash 2026 roadmap page now lists four Q4 deliverables with concrete completion dates. None shipped this week; what matters is that the team's published schedule has moved from quarterly buckets to fixed dates.
| AEP / Initiative | Completion | Category | Status |
|---|---|---|---|
| AEP-46 Spot Instances | (Draft since Dec 2024) | Standard/Interface | Draft |
| AEP-67 Console Bid Screening | Last Call since May 2025 | Standard/Core | Last Call |
| Custom Domain Configuration via Akash Console | Nov 15, 2026 | (TBD) | Estimated |
| Managed Services Market (AEP-11) | Dec 15, 2026 | (TBD) | Estimated |
| Onchain Provider Incentives (AEP-53) | Dec 31, 2026 | (TBD) | Estimated |
| Akash on Shared Security (AEP-79) | Dec 31, 2026 | (TBD) | Estimated |
What each tells you:
- AEP-46 Spot Instances has been in Draft for 23 months (Dec 1, 2024). Cosmos-SDK governance work appears to be the constraint; the AEP specifies a price-discount alternative to public-cloud spot pricing.
- AEP-67 Console Bid Screening has been in Last Call for 16 months (May 2025). The AEP-67 text notes the feature was renamed from "Bid PreCheck" to "Bid Screening" everywhere while the URL stayed. The Sep 1 weekly confirmed AEP-67 was delivered alongside AEP-46 on Aug 30/31.
- AEP-79 Shared Security is the security/economics pillar the Cosmos-to-Solana migration debate has been circling. Late-2026 target means a chain decision (or a no-migration defense) is structurally required by year-end.
- AEP-11 Managed Services Market (Databases-as-a-service) is the addressable-market-expansion play. The Sep 1 weekly noted AEP-46 + AEP-67 both delivered Aug 30/31; this week's confirmation is that those were the preceding gates to the Q4 deliverables.
- CMC AI Sep 7 AKT call: traders watching the tug-of-war between bullish structure ($2 target from a prominent analyst) and bearish momentum (sell pressure testing support). AKT traded ~$0.465–$0.561 across the Sept 4–7 window per CMC + cryptoRank.
Bittensor — Exploit Summit + AI Infra Summit Confirmed, GTAO Still in Window
Two Bittensor-relevant industry stages now bracket the September calendar:
| Event | Dates | Location | Notes |
|---|---|---|---|
| AI Infra Summit 2026 | Sep 15–17 | Santa Clara Convention Center | "Bottom four layers of Jensen Huang's AI layer cake" (data, models, compute, movement) |
| Exploit Summit 2026 (Bittensor's native conference) | Sep 28–29 | New City Gas, Montreal | "100+ subnet teams"; first speakers include Jon Durbin (Chutes / SN64) + Rob Myers (Manifold) |
What the lineup tells you:
- Exploit Summit 2026 speakers confirmed (Aug-prior livestream, Sep-this-window lineup): Jon Durbin is the core developer of Chutes / SN64, the decentralized serverless AI compute platform with TAO micropayments, GPU validation, DDoS protection, and TEEs; Rob Myers is Manifold's CEO. The "100+ subnet teams" framing aligns with the SN64 throughput lead from the Aug 25 weekly (Chutes at 425B tokens/day, more than the rest of the DePIN compute networks combined). The event is hosted by Bittensor Commons + Jose Rios + Etienne Leroy + Tensor Media Group.
- AI Infra Summit is the broader industry stage for the same week. Sponsors from hyperscaler, semiconductor, and data-center operators will be present; the "AI Data Center" track and "Physical AI" track are the two that intersect with decentralized compute claims.
- Grayscale GTAO S-1 decision window still open. No SEC ruling this week. Grayscale's Apr 3 S-1 to convert Grayscale Bittensor Trust to a spot ETF on NYSE Arca (ticker GTAO) is the procedural move that signals SEC engagement rather than dismissal. The Grayscale AI Fund retained 43% TAO allocation per the Sep 1 weekly. Bitwise TAO Strategy ETF (60% TAO + rest in TAO ETP) is in flight alongside the GTAO application.
io.net — IDE Burn Mechanism Holds, $8M Deal Context + Mint Immutability
Three observations about io.net's standing this week:
- IDE (Incentive Dynamic Engine) live since Jun 11, 2026 with first $IO burn the same day. Target: 12M IO burned over the first year (50%+ of post-payout network revenue). The CoinMarketCap Sep 7 brief frames IO as a small-cap asset (~$45M market cap) with real revenue and a deflationary mechanism — the cheapest small-cap narrative in DePIN compute.
- Mint-immutability confirmed (Own Your Mind independent RPC read):
mintAuthority = NULLon the IO SPL mint (BZLbGTNCSFfoth2GYDtwr7e4imWzpR5jqcUuGEwr646K, 8 decimals). Supply at 798,538,520.02 IO. Combined withfreezeAuthority = NULL, the SPL layer is immutable — only the IDE burn can move supply. - Pricing context: Aug 27 io.net blog post "The Token Cost Paradox: Why cheaper inference doesn't mean cheaper AI bills" is the public stance against the Gartner "AI compute costs are falling" thesis. Useful for understanding why io.net's burn mechanism is the explicit answer to the cost-collapse argument.
Prime Intellect — Post-INTELLECT-3 Research Cadence; Prime Agent ARC-AGI-3 Coverage Settled
The Prime Intellect repository (primeintellect-ai/prime) pushed on Sep 7, 2026 — 317★. What shipped in the public research log:
| Date | Item | Type |
|---|---|---|
| Aug 5 | Prime Agent released (95.0/95.2/95.5% across three runs; Best@3 99.97%; all 183 ARC-AGI-3 levels completed) | Announcement |
| Aug 13 | Prime Flash MoE — Blackwell-optimized MoE kernels | Research |
| Aug 14 | Measuring Autonomous AI Research | Research |
| Aug 25 | Uncovering a universal offline sandbox escape | Research |
| Sep 6 | Startup Fortune coverage of Prime Agent ARC-AGI-3 settled (95.5% Best@1 using Opus 5; surpassed reported human-expert baseline 95.4%) | Press |
The Sep 6 press coverage is the moment Prime Agent moves from "research result" to "stable benchmark achievement." AlphaSignal's coverage frames it as: gains showed up across models, not just one; the human-expert baseline was 95.4% as reported; all frontier models were below 1% when ARC-AGI-3 launched. The thesis is now: scaffolding design (the Prime Agent self-rewriting harness) buys you frontier-class benchmark performance without retraining.
Prime Intellect's stated team of backers: Founders Fund / Radical / NVIDIA / Intel / Andrej Karpathy / John Schulman / Dylan Patel / Clem Delangue. Customer stories (Zapier, Ramp) are the Aug 25 weekly material; the Sep window is research-cadence, not customer-cadence.
Marlin Oyster — EC2 Attestation SDK Surface
Two Marlin-side developments crossed the wire this week:
- EC2 instance attestation shipped in the Marlin Oyster SDK (Jul 15 Marlin X post, surfaced via marlin.org): the SDK now ships the verification side of EC2 attestation, meaning users of the SDK can check that a workload runs the expected code in a locked-down environment before handing it anything private. The checking logic stays upstream.
- Sep 4 Marlin X post: ongoing TEE-app guidance ("the TEE keeps everything sealed while it runs, the host just sees a black box, they can't open it, can't copy it, can't even-"). Marlin's developer-facing voice is the most operationally clear in the TEE-on-AWS space.
The x402-gateway work from the Sep 1 weekly (USDC payments across EVM + Solana + Oyster TEE signatures) is the protocol side; EC2 attestation is the platform side. They are the two halves of Marlin's "TEE coprocessors for decentralized AI" pitch.
Bacalhau — Sep 6 Push; NCL Protocol Now Default
The bacalhau-project/bacalhau repo pushed Sep 6, 2026 (871★). The Sep 1 weekly noted NCL Protocol merged as default; this week's push confirms it remains the canonical orchestration path for Compute-over-Data (CoD) jobs across the IPFS / Filecoin / S3 / edge surface. Expanso (commercial sponsor) runs the support business; the open-source repo is the upstream.
Golem — Aug 13 GLM Migration Invariant Holds
Own Your Mind's Aug 13 supply verification confirms the GLM migration invariant still binds: 796,565,516.94 GLM + 203,434,483.06 GNT = exactly 1,000,000,000.00. GLM is up ~195,712 since the June 2026 anchor and GNT down by the same amount — confirming one-to-one migration only, no new issuance. The Sep 1 weekly's Modelserve AI suspended + Blog frequency quarterly + Yagna velocity low are still the active state. The Salad.com GPU engineering trial (per CoinMarketCap Sep 7) is the only narrative-bright thread; whether S&P 500 customers end up paying in GLM is the open question.
Fluence — FLT/Stables Thread + Sep 6 Price Floor
Bybit Sep 6: FLT market cap $2.09M (–10% over 24h; price $0.00613762; 24h vol $130.05K; circulating 340.05M; max 1B). FLT ranks #2284 by market cap. The Sep 1 weekly noted Spheron GPU live + FLT-backed stablecoin as the active development surface; the Sep 6 read does not yet reflect that thread in price.
Fluence's position in the table:
| Network | Last-week tokens (OpenRouter) | Confidential-AI menu | Token price floor (Sep 6) |
|---|---|---|---|
| Phala | 24.0B (#5) | 10+ models, dstack-audited | (not measured) |
| AkashML | 23.6B (#6) | 6 models | $0.465–$0.561 |
| io.net | 71.45B (#2 DeAI Dashboard) | B200/H200-class | $0.45–$0.55 (Sep 7 CMC) |
| Fluence | n/a (sells compute, not inference) | Aqua/Marine WASM stack | $0.00614 |
Fluence's price is in a different regime — it sells WASM serverless compute, not inference-throughput tokens — but the network's existence is increasingly contingent on whether the Spheron partnership + FLT-backed stablecoin thread can convert to a measured utility story before the Q4 funding cadence matures.
Evidence at a Glance
| Project | This week's signal | Source | Cadence |
|---|---|---|---|
| Phala | 40.4B tokens/day peak (Sep 2 home-page snapshot); dstack audit cleared; 10 active models | phala.network | Weekly |
| Akash | Q4 2026 roadmap: 4 firm dates; AEP-46 Draft; AEP-67 Last Call | akash.network/roadmap/2026 | Quarterly cadence on roadmap |
| Aethir | ACCELERATE confirmed; IDC tokenomics = burn + variable fee; Coinbase ATH perp delist Sep 3 | aethir.com + ME News + ownyourmind.ai | Aug 24 launch, Sep 7 tokenomics framing |
| Bittensor | Exploit Summit Sept 28–29 lineup; AI Infra Summit Sept 15–17; GTAO still in window | exploitsummit.com + ai-infra-summit.com | Monthly |
| io.net | IDE live Jun 11; 12M IO burn target; mint immutable | globenewswire.com + ownyourmind.ai | Stable |
| Prime Intellect | Sep 6 ARC-AGI-3 press settled; Sep 7 repo push | startupfortune.com + GitHub | Weekly |
| Bacalhau | Sep 6 push; NCL Protocol default | GitHub | Steady |
| Marlin Oyster | EC2 attestation SDK; x402-gateway USDC | marlin.org | Steady |
| Fluence | FLT market $2.09M Sep 6; Spheron + FLT-stables in flight | bybit.com | Daily |
| Golem | Aug 13 GLM migration invariant | ownyourmind.ai | Quarterly |
What to Watch This Week
| Trigger | When | What it confirms |
|---|---|---|
| AI Infra Summit keynote coverage | Sep 15–17 | Whether decentralized compute networks are platform speakers, exhibit-floor sponsors, or absent |
| SEC GTAO decision (procedural) | Open window | Whether Grayscale's GTAO S-1 advances toward a ruling |
| Phala home-page next server-render | Rolling | Whether 40B/day holds in the rolling-window or 28B is the structural floor |
| Aethir tokenomics live parameters | Sep/Oct (per ownyourmind.ai: "no details yet") | Whether burn is fixed-rate, volume-tiered, or governance-tunable |
| Exploit Summit agenda refresh | Pre-Sep 28 | Whether subnets beyond Chutes/Manifold land on the speaker list |
| Akash Cosmos-to-Solana decision | Late 2026 | Whether AEP-79 Shared Security is migration-prep or independent path |
| Coinbase ATH spot decision | Open | Whether ATH remains listed on Coinbase post-Sep 3 perp delist |
Primary Sources
- Phala home-page server-render (Sep 2 = 40.4B, Sep 5 = 28.1B) — phala.network
- Phala/dstack audit announcement (Aug 13) — phala.com/posts/dstack-completes-security-audit-a-milestone-for-confidential-cloud
- dstack repo org rename (
Dstack-TEE/dstack) — github.com/Dstack-TEE/dstack - dstack-cloud archive (
Phala-Network/dstack-cloudarchived Jun 10) — github.com/Phala-Network/dstack-cloud - dstack Attestation Pipeline Hardening (Jun 16) — phala.com/tags/dstack
- Akash 2026 Roadmap — akash.network/roadmap/2026
- AEP-46 Spot Instances — akash.network/roadmap/aep-46
- AEP-67 Console Bid Screening — akash.network/roadmap/aep-67
- AEP-79 Shared Security — akash.network/roadmap/2026 (anchor)
- Aethir ACCELERATE — aethir.com/blog-posts/aethir-launches-accelerate
- Aethir tokenomics framing (Sep 7) — kucoin.com/news/flash/aethir-launches-accelerate
- Aethir ACCELERATE buildout analysis (Aug 24) — buttondown.com/ownyourmind/archive/aethir-hired-its-own-biggest-holder
- Exploit Summit 2026 — exploitsummit.com + luma.com/exploitsummit26
- AI Infra Summit 2026 — ai-infra-summit.com
- Grayscale GTAO S-1 — finance.yahoo.com/news/grayscale-files-bittensor-spot-etf-114048967.html + tradingview.com/news/coinpedia:861fc7055094b
- io.net IDE / 12M IO burn (Jun 11) — globenewswire.com/news-release/2026/06/11/3310232
- io.net tokenomics verification (Aug 13) — ownyourmind.ai/projects/io-net
- Prime Agent ARC-AGI-3 (Sep 6 coverage) — startupfortune.com/prime-intellects-prime-agent-just-beat-humans-on-the-arc-agi-3-benchmark + alphasignal.ai/news/prime-intellect-s-prime-agent-beats-human-experts-on-arc-agi-3-by-rewriting
- Prime Intellect research cadence — primeintellect.ai
- Marlin Oyster EC2 attestation SDK — marlin.org + Marlin X feed
- Bacalhau repo — github.com/bacalhau-project/bacalhau
- Golem GLM migration invariant — ownyourmind.ai/projects/golem
- Fluence FLT Sep 6 read — bybit.com/en/price/fluence-2
- AKT price (Sep 7) — coinmarketcap.com/cmc-ai/akash-network/latest-updates
- CMC AI IO small-cap framing (Sep 7) — coinmarketcap.com/cmc-ai/io-net/latest-updates
- GLM price/sentiment (Sep 7) — coinmarketcap.com/cmc-ai/golem-network-tokens/price-prediction