The Trust Stack · Weekly Signal Board

Proving it is the product.

AI capability is no longer scarce. What is scarce is the ability to prove what an AI did, why it did it, and that the resources behind it were legitimate.

$1T

Market capitalisation erased from AI chip and memory names in the last week of July 2026 - while several of those companies posted record profits. The market has begun pricing compute as a commodity whose returns must be demonstrated rather than assumed. See the Nvidia item below for the strongest rebuttal that read has yet drawn.

Evidence

The numbers that survive verification.

Every figure below was traced to a primary source. Anything that could not be re-traced this week was held at its prior value and date, or removed - never softened. See the integrity note at the foot of this page.

The stack

Five layers, and where the evidence is moving.

The framework is stable. What changes week to week is the strength of the evidence behind each layer. Direction reflects developments in the trailing quarter, not price action. No direction changed this week. The week's action landed in Layer 2 again, and it kept pushing the same way it has for a month - toward verification that comes from a party other than the one being verified. This week that arrived as product rather than principle: a security firm and OpenAI put out an inspector that audits other companies' agents before an enterprise trusts them, and a second security firm began certifying agents it did not build. Both strengthen Layer 2, so it holds at Strengthening. What none of it settles is still the whole game - whether the durable version of the proof is independent or absorbed into the platform an agent runs on. Verification and protocol strengthening, compute weakening, application and edge mixed.

What moved

This week, and the weeks behind it.

Independent verification stopped being a principle and started being a product this week. A security firm and OpenAI announced an inspector that audits other companies' agents - and their MCP servers - before an enterprise deploys them, with human researchers in the loop; days earlier a second security firm launched a certification for agents it did not build. Two of these in a fortnight is the independent-attestation side of this thesis starting to look like a market rather than an argument. I have kept the platform-native item below them on purpose, because it is the thing that side has to out-trust to earn a reason to exist, and the Nvidia and EU items behind it because the honest read needs the rebuttal and the regulatory clock in the same frame. Older items roll off to keep the board current, newest first. Where I could not trace a claim to a primary source in one hop, it is not here.

• 3 Sep 2026

  A third party starts auditing agents adversarially - the exact job a platform cannot credibly do on itself

  ▲ For

  Tenable and OpenAI announced the CyberAgents Exchange AI Inspector, which screens community-built agents, skills, MCP servers and multi-agent playbooks before an enterprise deploys them, pairing automated assessment with OpenAI's cyber models against review by Tenable's own researchers. My read: this is the independent-attestation half of the thesis showing up as something you can buy rather than something I assert. The whole case for a trust layer that survives a hyperscaler shipping its own is that some verification has to come from a party the builder does not control and is willing to be adversarial about it - and a security firm vetting someone else's agents, humans in the loop, is exactly that shape. Two honest limits I have to keep naming: it is announced as expected in September rather than live in my hands today, and one vendor's inspection is a service, not yet a standard. For on the direction.

  Tenable press release, announced at the OpenAI Cyber Summit, 3 Sep 2026; availability expected September 2026

• 31 Aug 2026

  CrowdStrike puts a certification behind agents it did not build - independent, with a catch I have to name

  ▲ For

  CrowdStrike launched an AI Partner Specialization whose Verified Agent certification validates partner-built agents against its requirements before they reach its marketplace, with Accenture, Anthropic and CoreWeave among the named partners. My read: this is the same move as the Tenable item - a party that did not write the agent standing behind whether it can be trusted - and two of these inside two weeks is what a market forming looks like, not a coincidence. The catch, and I will not bury it: CrowdStrike is certifying agents against its own bar, for its own marketplace, which is a real step toward independent but not the whole distance to a proof that a regulator or a court would accept from someone who trusts neither the builder nor the certifier. For on the direction, with that limit stapled to it.

  CrowdStrike newsroom, 31 Aug 2026 (AI Partner Specialization; Verified Agent certification)

• 26 Aug 2026

  Nvidia posts a record quarter and says "compute is revenue" - the sharpest rebuttal yet to the line I open this board with

  ◆ Mixed

  Nvidia reported second-quarter revenue of $96.2B, up 106% on the year, with data-centre revenue of $89.0B, up 117%, and guided the current quarter to about $108B. Jensen Huang's framing was "AI has reached its inflection point... now, compute is revenue" - which is close to a direct shot at the sentence at the top of this page, that the market has started pricing compute as a commodity whose returns must be demonstrated rather than assumed. So I have to mark my own homework: at the chip layer the returns are being demonstrated, loudly, and I am not going to bury that in a footnote. My read: two things are true at once. This thesis is about the accountability layer above compute, and a monster quarter for the biggest compute vendor does not touch it - and yet the shares still slipped after the beat, with the market fixed on financing loops and whether the enormous committed spend behind the boom converts to durable returns. Booming compute and doubted compute economics are not a contradiction. That unease is the same one this thesis is built on.

  NVIDIA Q2 FY2027 results, re-confirmed against the filing on sec.gov, 26 Aug 2026; after-hours market reaction reported 26 Aug 2026

• 22 Aug 2026

  Google Cloud finishes wiring agent identity into the platform itself - the verification primitive, shipped as an included feature

  ▼ Against

  Google Cloud brought its Agent Identity Auth Manager and Agent Identity APIs to general availability, completing a native agent-identity stack that sits on top of the cryptographic identity it has run since the spring: every agent gets a unique SPIFFE name and an X.509 certificate rotated daily, tied into the platform's own access policy and audit logs, in place of a shared static key. Read plainly, this is the primitive this thesis is built on - proof of who an agent is and a record of what it did - delivered as part of the cloud you already buy, not as a third-party product. My read: two things at once, and I will not bury either. It confirms the primitive is real and wanted, which is the thesis. And it is a platform vouching for the agents running inside it, which is self-certification - the bear case made concrete, and still the anchor of that case here. It is genuinely useful. It is also exactly the thing the independent efforts above it, from Tenable to CrowdStrike, have to be more trustworthy than to earn a reason to exist.

  Google Cloud Agent Identity documentation (primary); Auth Manager and APIs reached general availability 22 Aug 2026

• 2 Aug 2026

  The EU transparency clock started - the part that did not slip

  ▲ For

  Article 50 of the AI Act took effect: providers must mark synthetic audio, image, video and text in a machine-readable, detectable format, deployers must disclose deepfakes, and people must be told when they are talking to an AI or being read by emotion or biometric systems. Penalties reach 15 million euros or 3% of worldwide turnover. This is the counterweight to the deferral below, and I should keep saying it as loudly: the provenance-and-disclosure leg arrived on time. Machine-readable proof of what a model produced is now a legal requirement in Europe, not a nice-to-have. Systems already on the market get until 2 December 2026 to mark their output.

  European Commission transparency guidelines; in force 2 Aug 2026

• 27 Jul 2026

  The EU deferred the thesis's biggest tailwind by sixteen months

  ▼ Against

  Regulation (EU) 2026/1744 entered into force, moving the AI Act's high-risk obligations from 2 August 2026 to 2 December 2027, and embedded-product obligations to August 2028. Logging and traceability duties were postponed, not weakened - Article 12 still mandates lifetime event logging, with penalties to 3% of worldwide turnover. Any business case built on an August 2026 compliance deadline needs rebuilding. Held here as the counterweight to the item above: the disclosure leg is live, the audit-and-traceability leg is a 2027 event.

  Official Journal, 24 Jul 2026; in force 27 Jul 2026

Integrity note

What I changed this week, and what I held.

This week I added two items, rolled two off, and changed no numbers. The two additions are the same story from different vendors: Tenable and OpenAI announced an inspector that audits other companies' agents before deployment, and CrowdStrike launched a certification for agents it did not build. Together they are the independent-verification side of this thesis turning from a principle I argue into products you can buy, which is the most concrete support that side has drawn. I have left the Google Cloud platform-native item in place directly below them, because the honest way to show that progress is next to the thing it still has to out-trust, not instead of it.

I rolled two items off, and neither because it stopped mattering: the 1 September GLEIF discussion paper and the 24 August Okta Agent SSO item both made the independent-verification point, which this week's two shipping items now carry with commercial evidence rather than a proposal and a single-vendor login. I kept the Nvidia rebuttal, the Google Cloud bear case, and the two EU items so the board keeps a rebuttal and a regulatory clock in frame and does not tilt toward the reading that flatters the thesis.

On the numbers, I changed nothing, and I want to be plain about why. I could not re-reach the primary sources behind the six tiles this week - the fetch was gated, not the sources withdrawn - so under the verify-or-retain rule every figure stays at its prior value and its original source date rather than being refreshed on a date I did not confirm. No tile was contradicted this week. The one standing tension is unchanged: the hyperscaler capex tile still reads $650B, the Bridgewater figure from February, while newer trackers put big-four 2026 capex nearer $690-725B and one puts six-company spend far higher again - but those run through aggregators I could not trace in one hop, so the tile holds with the tension flagged rather than swapping in a number I have not confirmed. That rule is doing its job when it is inconvenient, which this week it was.

For context: the April edition of this thesis carried twenty-two market claims, of which seven were dropped and eight restated on re-verification - one figure overstated by roughly 170 times, token trading volume read as protocol revenue. The full correction log lives in the private edition.