Decentralized Compute Tuesday — October 06 2026 Weekly Update

Weekly covering September 30 – October 06, 2026. Four structural threads this week. First: Fluence shipped an MCP server on October 1, paired with x402 agentic payments — the first DePIN-compute platform to expose its full deploy/manage surface as a hosted Model Context Protocol endpoint (https://api.fluence.dev/mcp) for Cursor/Claude Code/VS Code/Codex/Gemini CLI/Copilot CLI/OpenCode/Cline/Kilo Code/Zed/Continue, with the agent able to pick a config, deploy it, read status, and shut it down from a plain-language prompt. x402 in tandem: an agent with a wallet sends USDC on Base to Fluence over plain HTTP, the amount lands on a Fluence balance in the same request, the paying wallet becomes the account, and the agent creates its own API key and plugs it into MCP — i.e. an agent that pays for its own compute and deploys it, start to finish, with no human in the loop. Second: Akash AEP-85 Buildpacks slipped its Sep 30 estimated-completion date with no console PR and no cnb-runtime image — the AEP is still Draft, no PR has landed on akash-network/console referencing BuildpacksConfig.tsx or useBuildpackLanguageDetection, no automatic-deployment-CICD-buildpacks-template/ has been added to akash-network/awesome-akash, and the spec page on akash.network/roadmap/aep-85 still shows Estimated Completion 9/30/2026. The structural backdrop: Cloud Native Buildpacks graduated from the CNCF on Aug 11, 2026 (per cncf.io/announcements/2026/08/11/) — i.e. the upstream spec AEP-85 depends on is now a CNCF top-level project, and buildpacks/pack had a release in August (v0.40.9, Aug 9). The DX gap the W40 weekly flagged is now both more solvable (CNB is graduated) and more visible (Fluence MCP went GA the next quarter). Third: Phala shipped Private AI Proxy 0.3.0 on Sep 30 — hardware-verified gate between coding agents (Claude Code / Codex / Cursor) and cloud AI, with DeepSeek Harness support for new web/desktop/headless/ACP sessions, agent-specific restart and session requirements surfaced in the app, and disconnect-revokes-local-token + restores-config semantics. This is Phala's third product of Q4 (after dstack-cloud v0.6.0 Sep 29 and the Sep 22 SayGM partner page) and the first one targeted at the AI-coding-agent persona rather than the GPU-tee-pod persona. Fourth: Bittensor Exploit Summit Sep 28-29 closed with a formal "thank you" on Oct 2 and a 25-sponsor roll-call on Oct 3 — the conference went live with Carbon winning Pitchtensor at 500 TAO, Minos demonstrating the second-incentive-mechanism agent research loop on a Mac with an iPhone pause-remote, Chutes' Jon Durbin explaining conviction-lock ("the acceleration, to an extent, makes us pause and decelerate") on the closing panel, and Steeves's Affine + Chutes + Yuma + 22 other ecosystem teams as sponsors. No V468 demo surfaced in the recap thread, but the conference is structurally the first Bittensor ecosystem gathering since the May 2026 EOC, and the "ecosystem in one room, arguing" framing is what Etienne Leroy (Opentensor Foundation) promised in the Sep 7 GFdaily interview. Below: io.net Agent Cloud is now in its third week of public messaging (Oct 3 X post, Jul 25 launch — MCP-native compute, sub-3-min provisioning, pay-for-work billing via Stripe + x402), Render Foundation Node Bonus Program accrual opened Oct 1 with claims opening Oct 8, Aethir DefiLlama Sep 30 read shows $4.57M fees 30d / $914,873 revenue 30d / $183.21M cumulative fees / $36.64M cumulative revenue (down from Q3 2025's $39.8M quarterly record), and the Sep 22 "Q3 2026 transparency report still open" trigger is now closed (Q4 2026 delivery).

Macro Thread — Fluence MCP Server: First DePIN-Compute Platform With Hosted MCP Endpoint + x402 Agentic Payments

The Sep 22 weekly covered Fluence VPC (private networking for VMs) as the first "non-GPU" infrastructure release of 2026. The Oct 6 reading is a second-order claim: Fluence is now the first DePIN-compute network with a hosted MCP server (per the Oct 1 fluence.ai/blog/fluence-mcp-server post by the team), and the first to pair MCP with x402 for agent-paid compute. The structural shift: the previous DePIN-compute UX required opening a Console, picking a config, signing transactions; the Oct 1 reading replaces all of that with "tell your agent what you need in plain language."

From fluence.ai/blog/fluence-mcp-server (Oct 1, 6-min read):

What the table tells you about the Oct 1 release:

1. MCP is the new "Console" for AI-agent customers. The previous DePIN-compute UX required the agent's human owner to open a Console, navigate the deploy flow, and monitor the deployment manually. The Oct 1 release replaces that with: agent gets a config snippet, agent reads the available actions via MCP, agent calls deploy(config_name) from a prompt, agent calls status(deployment_id) to verify, agent calls shutdown(deployment_id) to clean up. The console is now an internal detail of the agent's tool surface, not a human-facing UX.
2. x402 as the agent-payment primitive is what makes MCP + no-account possible. The MCP server alone would still require the user to fund a Fluence balance with USDC via the Console. The x402 flow removes that: agent sends USDC on Base to Fluence, the amount lands on the balance in the same request, the wallet that paid becomes the account, and the agent creates its own API key with the wallet's signature as the auth primitive. This is the first DePIN-compute platform where an agent can be deployed, pay for its own compute, and shut down — with no human in the loop and no email signup.
3. The 11-client config list is the AI-coding-assistant footprint. Cursor + Claude Code + VS Code + Codex + Gemini CLI + Copilot CLI + OpenCode + Cline + Kilo Code + Zed + Continue is the canonical "every AI coding tool that matters in Q4 2026" list. The Fluence Console's MCP page generates per-client config snippets (each one names the file the client reads from: .mcp.json for Claude Code, settings file for Cursor, etc.). This is the kind of "first-class DX" pattern Phala dstack-cloud v0.6.0 shipped for the CVM-Console path (Sep 29) and Marlin Oyster's x402-gateway shipped for the inference-payment path (Sep 1).
4. The model of "agent holds API key, acts on user's account, spends user's balance" is the same model as Phala's agent-wallet pattern. Phala's solutions/ai-agents page shows a sign-RPC action log of cal-bot, pnl-monitor, inbox-triage, research-bot, devops-bot, support-bot-v3.2, cal-bot calling calendar.read, wallet.readbalance, gmail.composedraft, web.search, github.writepr.merge, zendesk.writeticket.reply, calendar.writecreate with action-level signed RPCs. Fluence's MCP server is the compute-deploy equivalent: agent calls fluence.deploy, fluence.status, fluence.shutdown with the API key as the auth primitive. Both patterns are doing the same thing — making the agent the operator, with the user's key/balance as the trust anchor.
5. The Claude Code config lives in .mcp.json (project root) but reads the key from a shell env var. The blog post quotes the exact config: {"mcpServers": {"fluence": {"type": "http", "url": "https://api.fluence.dev/mcp", "headers": {"X-API-KEY": "${FLUENCE_API_KEY}"}}}}. This is the canonical "don't put secrets in the repo" pattern for coding-agent tool config. The other 10 clients have user-level configs (Cursor's ~/.cursor/mcp.json, etc.), which is the same defense-in-depth pattern.
6. The blog explicitly disclaims the danger: "An agent holding your API key acts on your account and spends your balance. Three things keep that predictable: scope the key, keep the key out of your repo, ask for the price first." This is the same threat model as every other "agent holds a key" pattern (Anthropic's claude.ai, OpenAI's Codex CLI, Google's Gemini CLI) and Fluence is naming it explicitly rather than glossing over it. The "ask for the price first" guard is the agent-deploy-specific part — it forces the agent to surface the price before the deploy is signed.

What MCP + x402 Means for the DePIN-Compute Sector

Three observations on the structural shift:

1. The "Console → Agent" transition is now an open race. Fluence shipped MCP + x402 on Oct 1. Phala's agent-primitive is the sign-RPC action log (CVMs hosting Claude Code/Codex/Vijil, sealed by the compose-hash, with every diff signed before merge). Render Network's OTOY integration is the 3D-rendering tool chain. io.net's Agent Cloud (Jul 25, Oct 3 re-promoted) is MCP-native compute with sub-3-min provisioning, pay-for-work billing via Stripe + x402 — a parallel implementation to Fluence. The four networks are now in different shapes on the agent-primitive axis: Fluence = hosted-MCP + x402 (Oct 1), io.net = Agent Cloud + Stripe + x402 (Jul 25), Phala = sign-RPC + sealed-OAuth (Q3-Q4 2026), Render = OTOY plugin chain (long-standing). The first three are explicitly built for AI agents; Render is built for human 3D artists with AI tooling.
2. The MCP server being hosted (not self-hosted) is the key developer-experience choice. A self-hosted MCP server would require the user to install and maintain the server binary; the hosted model means the user only has to add the config snippet and restart their client. This is the same trade-off as "Console is a hosted web app" vs "deploy your own console from source" — the hosted version has lower friction at the cost of vendor lock-in. For the early-agent-CLI market, hosted wins on friction.
3. x402 + USDC on Base is the cross-network payment primitive. x402 (HTTP 402 Payment Required) is the standard that Coinbase, Cloudflare, and others have been pushing since 2024. Fluence's x402 implementation accepts USDC on Base, lands the amount on a Fluence balance in the same request, and uses the paying wallet as the account. Gonka's x402 facilitator (Sep 18, first live GNK settle) is a parallel implementation on a different chain; Marlin Oyster's x402-gateway (Sep 1) is a third. The convergence on x402 across Fluence / Gonka / Marlin means a wallet-based AI agent can pick any of the three networks, pay for compute, and get the same auth primitive. This is the closest thing to a "DePIN-compute payment standard" that has shipped in 2026.

What Fluence Hasn't Shipped (Yet) — The Open Items

Two things the Oct 6 reading can't confirm:

1. No public adoption metrics yet. The Oct 1 blog post doesn't quote any usage numbers (active agents, deploys, x402 settles, balance load). The Sep 8 VPC release also didn't ship with adoption metrics. The Sep 22 weekly flagged "no Sep 22-specific Fluence news" and the Oct 6 reading adds the same gap. Fluence is now 3 weeks into the post-VPC period and 6 days into the post-MCP period; the first public "agents deployed via MCP" count is the next adoption signal to watch.
2. The hosting location of the MCP server (api.fluence.dev/mcp) is not specified. It could be on Fluence's own cloud (Fluence runs on Arbitrum Orbit L2 + its own provider network), on a third-party (Cloudflare Workers, Fly.io), or on the same hosted infra that runs the Console. The Oct 1 blog doesn't disambiguate. This matters for trust-minimization: a hosted MCP server on a centralized cloud is a single point of failure for the agent-UX primitive; a hosted MCP server on Fluence's own provider network is a proof-of-dePIN primitive (the network is being used to host the network's own DX).

Phala — Private AI Proxy 0.3.0 (Sep 30) Ships the Coding-Agent Persona

The Sep 22 weekly covered Phala's dstack-cloud v0.6.0 (the CVM persona) and the SayGM partner page (the inference-gateway persona). The Sep 30 reading is Private AI Proxy 0.3.0 — the third product line, targeted at the AI-coding-agent persona (Claude Code / Codex / Cursor).

From the phala.com/posts/private-ai-proxy-hardware-verified-private-ai-for-coding-agents (Sep 30, 5-min read):

What the table tells you about the Sep 30 release:

1. Private AI Proxy is the "agent ↔ AI service" trust primitive. The threat model: a Claude Code / Codex / Cursor agent can read the user's entire codebase, and the AI service (Anthropic / OpenAI / Google) is the egress point for that code. If the AI service is compromised (a supply-chain attack, a malicious plugin, a model-prompt-injection leak), the codebase leaves the user's machine. Private AI Proxy puts a hardware-verified gate between the agent and the cloud AI service — the AI service is verified by a TDX/SEV-SNP/Nitro attestation before the proxy forwards the request, and the agent can refuse to talk to an unverified service.
2. The Sep 17 Plugin4Shell disclosure is the structural backdrop. Per shattered.io's Plugin4Shell analysis: "Plugin4Shell is a zero-click remote code execution vulnerability disclosed by Air Security on September 17, 2026. It affects the plugin installation process in four AI coding agents, allowing an attacker who controls a plugin's Git repository to bypass SHA-pinning and swap trusted code for malicious code." The four affected agents: Claude Code, Codex, GitHub Copilot, Gemini CLI. The Plugin4Shell disclosure is the first "AI-coding-agent as supply-chain attack surface" story of 2026, and Phala's Sep 30 Private AI Proxy 0.3.0 release is positioned as the answer: even if the plugin is compromised, the proxy verifies the AI service by attestation, so the malicious plugin can't talk to an unverified AI service.
3. The three Phala personas now cover the three trust-boundary cases. (a) dstack / dstack-cloud / Confidential AI Inference — the "model is in the TEE" persona; (b) SayGM / OLLM / Private AI Gateway — the "inference gateway with verifiable privacy" persona; (c) Private AI Proxy 0.3.0 — the "agent ↔ AI service trust" persona. The three are complementary, not competitive: a developer using dstack for confidential inference can still use Private AI Proxy to gate the agent that calls the dstack endpoint.
4. The disconnect semantics are the canonical "restore my config" pattern. Disconnect → revoke the local token → restore the config the proxy took over → user edits outside the app are preserved. This is the same UX pattern as Phala's CVM-disconnect (terminate the lease, return the resources, preserve the local file system) and the same trust pattern as Phala's KMS key release (RA-TLS binds the workload to its attestation, so the key is only released while the workload is verified).

Why the Persona Split Matters

Three observations on the structural shift:

1. Phala is now the only DePIN-compute platform with three named personas for the three AI-customer segments. dstack = "I run a model in a TEE." SayGM / OLLM = "I want verifiable private inference on someone else's model." Private AI Proxy 0.3.0 = "I want my agent to talk to a verified AI service." Each persona is a different trust boundary, and each one has a different customer. The persona split is what lets Phala address the developer-who-runs-models, the enterprise-that-buys-inference, and the agent-builder in parallel.
2. The "hardware-verified gate" framing is the same trust primitive as dstack's RA-TLS, but applied to a different boundary. dstack's RA-TLS binds the workload to its attestation so the KMS only releases keys to the verified workload. Private AI Proxy 0.3.0 binds the AI service to its attestation so the agent only forwards requests to the verified service. Both are "verify the other side by hardware attestation before the trust-bearing action"; the difference is the direction of the trust-bearing action (KMS release of key vs. agent release of request).
3. The Sep 30 ship date is one day after dstack-cloud v0.6.0 (Sep 29). The two releases are back-to-back, which is consistent with the Q4 2026 Phala product cadence (a major release every 2-3 weeks per the Aug 4 weekly's product observation). The next Phala cadence checkpoint is in the Oct 13-20 window per the prior pattern.

Akash — AEP-85 Buildpacks Slips Sep 30 Estimated-Completion, No Console PR, CNCF Upstream Is Now Graduated

The Sep 22 weekly flagged AEP-85 Buildpacks as "Whether the developer-experience gap with Phala closes" — the W41 reading is that the Sep 30 estimated-completion date was missed with no PR on akash-network/console and no cnb-runtime image published to ghcr.io/akash-network/. The structural backdrop is the CNCF graduation of Cloud Native Buildpacks on Aug 11, 2026, which means the upstream spec AEP-85 depends on is now a top-level CNCF project.

From the live akash.network/roadmap/aep-85 page (still Draft as of Oct 6):

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

1. The Sep 30 deadline was missed by T+6 days (Oct 6) with zero console-side work in the pipeline. No PR on akash-network/console, no template on akash-network/awesome-akash, no image on ghcr.io/akash-network/. This is the second time in 2026 an Akash AEP with an estimated-completion date has slipped the date with no PR (the first was AEP-67 in late July, which got pushed by 2 weeks and then delivered Aug 30/31).
2. The CNCF graduation of Cloud Native Buildpacks on Aug 11, 2026 is the structural positive. Per the CNCF announcement: "CNCF® today announced the graduation of Cloud Native Buildpacks, an open source project that builds OCI-compliant container images directly from application source code. The graduation reflects the project's maturity in production adoption, vendor-neutral governance and security practices." This is the formal "spec is stable" signal for the upstream AEP-85 depends on. Akash adopting a CNCF-graduated spec is the right architectural choice; the implementation just hasn't happened.
3. buildpacks/pack v0.40.9 (Aug 9) is the latest upstream pack release. Per the pack releases page, the cadence is roughly 1 release per 2-3 weeks; v0.40.9 is the Aug 9 release. The CNB ecosystem is healthy: Heroku, Google Cloud Run, Fly.io, Cloud Foundry, VMware Tanzu all use CNB in production; the buildpacks.io spec is stable. Akash adopting CNB would put Akash's "Git-to-deploy" flow on the same primitive as the rest of the cloud-native ecosystem.
4. The DX gap with Phala dstack-cloud v0.6.0 is now both more solvable and more visible. Phala shipped the dstack-cloud CLI on Sep 29 (1.5 weeks before the AEP-85 deadline). Fluence shipped the MCP server on Oct 1 (the day after the deadline). Akash is the only one of the three networks whose "developer builds and deploys a containerized app from source" flow is still Draft. The Sep 22 weekly's "Phala dstack-cloud v0.6.0 closes the DX gap" reading is now confirmed: Phala shipped it, Fluence shipped the agent-front-door version, Akash is still drafting.
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 structural signal to watch is the first PR referencing BuildpacksConfig.tsx or cnb-runtime — that's the "implementation has started" marker.

What the CNCF Graduation Means for the DePIN-Compute Sector

Two observations:

1. The "buildpacks over Dockerfile" primitive is now a CNCF standard, not a vendor opinion. Every DePIN-compute network that wants a "Git-to-deploy" flow now has the same upstream spec to point to. Akash (AEP-85), any future Akash competitor (Marin's Render RNP series, the Phala dstack repo's own buildpack integration), and the wider cloud-native ecosystem (Cloud Run, Fly.io, Heroku) all converge on CNB. The risk for Akash is that competitors who adopt CNB faster will attract the "I want Heroku-style DX on a decentralized cloud" customer first.
2. The CNCF graduation also opens the door to vendor-agnostic CNB builders. The AEP-85 spec names paketobuildpacks/builder-jammy-base and heroku/builder:24 as the two vendor-published builder images. With CNB now a CNCF top-level project, more vendor builders (Google's gcr.io/buildpacks, Microsoft's mcr.microsoft.com/oryx/builder, etc.) become fair game as Akash's "BUILDPACKS_TEMPLATE_ID" choices. This is the same "open up the vendor surface" pattern as the AEP-85 AEP itself.

Bittensor — Exploit Summit Sep 28-29 Closed, "Thank You" Oct 2, 25-Sponsor Roll-Call Oct 3

The Sep 22 weekly noted Exploit Summit as "the conference V468 lands the same week as" and the Sep 29 weekly covered it as "in session." The Oct 6 reading is the post-conference wrap-up: the Oct 2 "That's a wrap!" post and the Oct 3 sponsor roll-call.

From the Exploit Summit @ExploitSummit X feed (Oct 2-3):

What the table tells you about the W41 Exploit Summit wrap:

1. The conference happened and the ecosystem showed up — but V468 was not the centerpiece. The Sep 22 weekly's framing was "V468 mainnet live + Exploit Summit as the live-demo venue for basket trading." The actual Day 2 recap thread covered: Gamma proposal (service credits), Pitchtensor (Carbon at 500 TAO), Minos agent research on a Mac, conviction lock (Chutes' Durbin), and a closing panel. V468 is the substrate for all of this (every subnet running on mainnet is on V468 now), but the live demo surface was the subnet-level work, not the root-validator basket-trading work. This is consistent with the Exploit format: "ecosystem in one room, arguing," not "core dev shows off the new feature."
2. The 25-sponsor roll-call is the structural takeaway. Affine (Platinum) and Chutes (Platinum) are the two ecosystem anchors; the 23 other sponsors span AI infrastructure (Lium, KubeTEE, Score, BigTensor, ChronoLLM, Dendrite, Leadpoet, Yanez AI), DePIN (Green Compute, OpenRoboto, Gradients), tooling (Talisman, Yuma, 404gen, Bitcast, BitMind, Crucible Labs, Mentat Minds, BitstarterAI, HackConcordia, McGill AI Lab, SubnetsForGood, Vida.io, btlabs.ai), and the Foundation/Commons split (Opentensor Foundation, Proof of Talk, Manifold). This is the densest sponsor concentration in the Bittensor ecosystem of 2026 — the conference was structurally a "show the network is real" event.
3. The Sep 26 TAO $315.67 / $3.58B mcap / +6.42% 24h was the pre-conference read. The Sep 26 reading is from CoinStats and lines up with the Sep 25-26 "TAO is up" coverage. The Oct 6 reading (no specific X tracker) is the structural follow-up: post-conference, no fresh TAO price action has been reported in the major outlets, and the "institutional inflows into Bittensor ecosystem reached $620M by early 2026" framing (per avark.agency) holds.
4. The Sep 22 weekly's open Exploit-related triggers are now closed. "Whether V468 basket-trading gets a live demo" — no V468-specific demo in the recap (the demos were subnet-level). "Steeves (Affine) and Durbin (Chutes) speakers" — both delivered (Steeves via the Affine sponsorship, Durbin via the closing panel). "Foundation/Commons split" — confirmed (Opentensor Foundation is a separate entity from Bittensor Commons, which produced the conference). "100+ subnet teams" — confirmed by the sponsor roll-call (the 25 sponsors are a subset of the 100+ subnets).

io.net — Agent Cloud Re-Promoted (Oct 3 X Post), IDE Burn On-Chain @ ~5.8M IO Purchased / 2.15M Burned

The Sep 29 weekly noted io.net's 4,562 cluster-ready GPUs/CPUs at 47.7% utilization and the Oct 11 next-unlock ($1.20M, 8.58M IO). The Oct 6 reading adds:

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

1. Agent Cloud is now in its third week of public messaging. The Jul 25 launch is the original date; the Oct 3 @ionet X post is the re-promotion. The product surface is: agent picks a GPU, agent picks a container, agent pays via Stripe or x402, agent gets billed only for completed work, agent shuts down automatically. This is a parallel implementation to Fluence MCP (Oct 1) and Marlin Oyster's x402-gateway (Sep 1) — same "agent-as-customer" primitive, different stacks.
2. The "agent builds code → tests it → picks a GPU → deploys container → pays → shuts down" framing is the Oct 3 reframing. Per the @ionet Oct 3 post: "Your agent writes the code. Then it tests it, picks a GPU, deploys the container, pays for it, and shuts it down. Nobody had to file a ticket or get woken up at 2am." The "no 2am wake-up" line is the structural contrast to the traditional "human on-call engineer" cloud model. Compare to Fluence MCP's "tell your agent what you need in plain language" framing.
3. The IDE burn pace reconciles on-chain. Per ownyourmind.ai's Sep 24 read: ~5.8M IO purchased by clients vs ~2.15M burned; the 294 epochs checked on Sep 19 each reconcile to the burn on the IO mint. The June reading of "emission-funded" is dated; the September reading is "purchases now exceed the burn on io.net's accounting." The purchases themselves settle partly in fiat and USDC off-chain, so they remain io.net's claim with an on-chain floor.
4. The Oct 11 next unlock (8.58M IO, $1.20M) is 5 days out as of Oct 6. Per CoinLaunch's vesting schedule. The Sep 11 unlock (14.14M IO, $1.87M at the time) is now 25 days in the rear-view. The Oct 11 unlock is the smaller of the two — the post-Sep-11 unlock overhang step-down is the empirical question.
5. No Oct 1-5 io.net blog post. The Aug 18 "What happens when you can't get a GPU? The hidden cost of cloud wait times" essay is the most-recent io.net blog post; the Jul 25 Agent Cloud launch is the most-recent product post. The Sep 1 → Oct 5 blog gap is the same pattern the Sep 22 weekly flagged ("the Aug 13 → Sep 1 cluster has gone quiet"). The next product post is the open question.

Aethir — DefiLlama Sep 30 Read: $4.57M Fees 30d / $914,873 Revenue 30d / $183.21M Cumulative Fees

The Sep 22 weekly flagged "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 Oct 6 reading is the DefiLlama live read plus the Q3 2026 timing update.

From DefiLlama's Aethir page (Oct 6 read):

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

1. The 30-day DefiLlama revenue of $914,873 vs Aethir's Q3 2025 marketing claim of $166M ARR is a ~14x gap. DefiLlama computes annualized revenue as (Revenue 30d) × 12 = $10.98M annualized. Aethir's Q3 2025 record (per aethir.com/blog-posts/aethirs-record-breaking-q3) is "$166M ARR" ($39.8M Q3 booked network revenue, 13x growth from $12M, 6.1% Rev/MC). The 14x gap is because (a) DefiLlama only captures 20% protocol fees (per DefiLlama's methodology note), while Aethir's "Q3 booked network revenue" includes total service fees (the full gross), and (b) the Q3 2025 figure is from a year ago, before the Q4 2025 → Q1 2026 → Q2 2026 → Q3 2026 trajectory of ~30-40% revenue contraction. The structural truth: Aethir's Q3 2025 $39.8M was a peak; the current run-rate is ~$11M annualized.
2. The Q3 2026 transparency report is structurally delayed to Q4 2026. The Sep 22 weekly flagged Q3 2026 transparency report (Q4 2026 delivery) as the open trigger. The Oct 6 reading confirms Q3 2026 closed on Sep 30 with no public Q3 2026 revenue number yet. The Q3 2026 report will likely land in late October or early November, after the Sep 30 quarter-end.
3. ATH circulating supply is 23.31B (Sep 16 read) = 9.0% of 42B cap; the Sep 22 weekly's "766M/month emission" pace continues. The Sep 16 reading is the most-recent; no new ATH unlock has been reported. The 23.31B / 9.0% / 42B-cap structure is the canonical "post-TGE, pre-second-vesting-cliff" Aethir state.
4. The ACCELERATE Phase 2 (20 MW US/EU buildout) is still no-update since the Aug 25 weekly. No new press releases or blog posts about the 20 MW buildout, the B300/GB300 procurement, or the Axe Compute $317M prepayments / $3.2B signed have surfaced since Aug 25. The Sep 22 weekly flagged the same gap. The W41 read: the ACCELERATE story is steady-state, not active news.

Caveat: DefiLlama Revenue Is Protocol-Fee Only

The 14x gap between DefiLlama's $914,873 × 12 = $10.98M annualized and Aethir's Q3 2025 $39.8M / $166M ARR marketing is structural, not contradictory. Per DefiLlama's methodology: "Revenue: 20% protocol fees charged by aethir." Aethir's $39.8M Q3 2025 figure is the total service fees (the full gross), not the protocol's 20% cut. Aethir's actual protocol revenue (the 20%) at the Q3 2025 peak would have been ~$8M/quarter / $32M annualized, which is closer to (but still above) the current $11M annualized read. The Sep 22 weekly's "6.1% Rev/MC ratio" test is the relevant test: Q3 2025 $39.8M / Q3 2025 mcap ~$650M = 6.1%; current $10.98M annualized / $122.62M current mcap = 8.95% (DefiLlama's compute) or, if we trust the Aethir-side gross, current annualized gross / current mcap = ~70%. The "6.1% test" reads differently depending on which revenue figure you trust.

Render — Node Bonus Program Accrual Opens Oct 1, Claims Open Oct 8; On-Chain Burn Continues

The Sep 22 weekly flagged "Render Node Bonus Program accrual Oct 1, claims Oct 8" as a watch-this-week trigger. The Oct 6 reading is that the Sep 16 → Oct 1 → Oct 8 timeline is now in the live window.

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

1. The Sep 16 → Oct 1 → Oct 8 → Oct 25 timeline is now in the "Oct 1 accrual live" phase. Node operators are now accruing bonus RENDER for past-frames-rendered; the first claims window opens Oct 8 (T+2 days from Oct 6). The Sep 22 weekly's "whether the Sep 16 announcement translates to net-new GPU onboarding" trigger is the empirical test for the next 7-14 days.
2. The Sep 25 small-GPU onboarding round is still open (Typeform app). The "fewer than 4 GPUs, 16GB+ VRAM, weekday availability" target demographic is the first Render onboarding round that names a specific operator archetype. The Sep 22 weekly flagged this; no Oct 1-5 update.
3. No on-chain burn tracker update for Oct 1-5. The Sep 16 reading (4,588 RENDER / $6,212 / 27.97% of daily emissions) is still the most-recent. The Sep 22 weekly flagged "the burn fraction, not the burn dollar amount, is the new headline metric for RENDER" — that framing holds; the next burn-tracker tweet will be the structural read.
4. The 5,600 total nodes / 81M total frames figures are cumulative-since-inception, not current. Per CoinStats' "October 2026 Analysis" — "However, these are cumulative figures. Current active users, active nodes, GPU utilization, customer retention, and recent job volume are not consistently disclosed." This is the same gap the Sep 22 weekly flagged on the supply-side transparency dimension; the Sep 16 Node Bonus Program is the supply-side push to expand the active-node count, and the Oct 8 claims open is the first test of whether the active count actually grew.
5. No Render blog post since Aug 2026. The Aug 2026 Monthly Report is the most-recent formal Render Network post. The Sep 16 Node Bonus Program + Sep 25 onboarding round are the only Sep-Oct announcements; both are supply-side pushes, not product launches.

Other Projects — Quick Status (Oct 6)

Note on Golem Network vs Golem Cloud

The Sep 22 weekly covered Golem (the GLM-token decentralized compute network, originally launched 2016, yagna as the Rust implementation, Modelserve AI suspended, 796.57M GLM + 203.43M GNT = 1B token invariant). The W41 reading surfaces a separate project: Golem Cloud (Golem Cloud Inc., John A. De Goes, "the agent runtime" — a WASM-based cloud platform with Golem 1.5 release on May 8, 2026, Golem 1.6 expected Oct 13, and the Sep 29 blog post "Making decisions with Jev in Golem"). The two projects share a name and a lineage (the original Golem Factory split in 2019) but are organizationally separate: Golem Network (GLM token) is the decentralized compute marketplace; Golem Cloud is a SaaS agent-runtime platform. The Sep 22 weekly's "Golem" entries refer to Golem Network; the Sep 29 Golem Cloud blog post is a separate project that does not affect GLM-token metrics. The Oct 6 reading flags this distinction because the blog.golem.cloud Sep 29 post can be confused with the Golem Network if a reader doesn't disambiguate.

What to Watch This Week

Primary Sources

• Fluence MCP server launch (Oct 1) — fluence.ai/blog/fluence-mcp-server
• Fluence x402 agentic payments — fluence.ai/blog/x402-agentic-payments/
• Fluence news index — fluence.ai/blog/category/news
• Fluence blog index — fluence.ai/blog
• Phala Private AI Proxy 0.3.0 (Sep 30) — phala.com/posts/private-ai-proxy-hardware-verified-private-ai-for-coding-agents
• Phala SayGM partner page (Sep 22) — phala.com/tags/AI
• Phala Private AI Agents (sign-RPC action log) — phala.com/solutions/ai-agents
• Phala dstack-cloud v0.6.0 release notes — docs.phala.com/dstack-cloud/release-notes
• 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 2026 Roadmap — akash.network/roadmap/2026
• Akash Network homepage (Oct 6 read: 61 providers / 422 GPUs / 86 TB / 731 TB) — akash.network
• CNCF Cloud Native Buildpacks graduation (Aug 11) — cncf.io/announcements/2026/08/11/cncf-announces-graduation-of-cloud-native-buildpacks-advancing-the-standard-for-container-builds
• Cloud Native Buildpacks v0.6.11 release notes (Jul 21) — techdocs.broadcom.com/us/en/vmware-tanzu/platform/elastic-application-runtime/10-2/eart/cnb-docs-cnb-rel-notes.html
• Bittensor Exploit Summit wrap (Oct 2) — x.com/ExploitSummit/status/2106063469866881384
• Bittensor Exploit Summit sponsor roll-call (Oct 3) — x.com/ExploitSummit/status/2106378324276601031
• Bittensor Exploit Summit Day 2 recap (Sep 29) — x.com/ExploitSummit/status/2105065369874686425
• Exploit Summit Luma — luma.com/exploitsummit26
• Exploit Summit Coinpedia — events.coinpedia.org/exploit-summit-2026-8817
• Bittensor TAO Sep 26 read ($315.67 / $3.58B / +6.42%) — coinstats.app/news/711c1e8d8fd81b3eaf5779f01ca39f472bf31babb47910819a87fa66af1a1856_Bittensor-TAO-News-Today-Why-TAO-Is-Up--26-September-2026
• How to build on Bittensor in 2026 (avark.agency) — avark.agency/learn/how-to-build-on-bittensor-in-2026-the-complete-guide-to-launching-on-the-decentralized-ai-network
• Is Bittensor (TAO) a Good Investment? October 2026 Analysis — coinstats.app/ai/a/investment-analysis-bittensor
• Aethir DefiLlama (Oct 6 read) — defillama.com/protocol/aethir
• Aethir Q3 2025 record ($166M ARR / $39.8M rev / 6.1% Rev/MC) — aethir.com/blog-posts/aethirs-record-breaking-q3
• Aethir 12-Month Strategic Roadmap — aethir.com/blog-posts/aethirs-12-month-strategic-roadmap-supercharging-enterprise-ai-compute-growth
• 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 Sep 17 +3% rally — coinmarketcap.com/top-stories/6aab3d5884f8f4387e494554
• Render Sep 21 +8% rally — coinmarketcap.com/top-stories/6ab18653d28b6a241fa1825f
• Is Render a Good Investment? October 2026 Analysis — coinstats.app/ai/a/investment-analysis-render-token
• io.net cluster dashboard Sep 29 — cloud.io.net/explorer/devices
• io.net explorer Sep 29 — explorer.io.net/
• io.net Agent Cloud re-promotion (Oct 3) — x.com/ionet
• io.net Agent Cloud launch (Jul 25) — tradingview.com/news/coinmarketcal:64a5cb3ce094b:0-io-net-launches-agent-cloud-25-jul-2026
• io.net IDE burn (Sep 24) — ownyourmind.ai/projects/io-net
• io.net IDE Goes Live (Jun 11) — globenewswire.com/news-release/2026/06/11/3310232/0/en/io-net-to-burn-up-to-12m-tokens-as-network-closes-8m-deal-and-hits-4-billion-daily-ai-tokens.html
• io.net vesting schedule — coinlaunch.space/projects/ionet/vesting
• Golem Cloud Sep 29 blog post — golem.cloud/blog
• Golem Network tokenomics review — ownyourmind.ai/projects/golem
• Plugin4Shell (Sep 17, 4 AI coding agents) — shattered.io/plugin4shell-ai-coding-agent-sha-pinning-bypass-2026
• Microsoft Reactor MCP events Oct 6 + Oct 8 — developer.microsoft.com/en-us/reactor/events/27243
• MCP 2026 roadmap (governance maturation under Linux Foundation) — toloka.ai/blog/the-future-of-mcp-enterprise-adoption
• 24-Hour Crypto Market Update Oct 3 — cryptorank.io/news/feed/7b27e-24-hour-crypto-market-update-what-happened-in-crypto-on-october-3-2026