The investment thesis
AI capability is no longer scarce. Proof is.
Raw model performance is converging and commoditizing. The durable value is accumulating one layer above it - in the systems that make AI outputs accountable, auditable, and defensible to a regulator, an insurer, a court, or a counterparty.
I call it the Trust Stack. This page is the thesis in public. It is republished weekly, including the weeks it gets weaker.
The argument
Regulated industries are not slow AI adopters. They are blocked ones.
The thesis rests on a structural asymmetry rather than on enthusiasm. Every serious buyer of AI - a hospital, a bank, a law firm, a government, an insurer - faces the same blocker, and it is not capability. It is that they cannot deploy a system whose decisions they cannot account for. The unblock is worth more than the model.
The second half of the argument is about agents specifically. As AI systems begin executing rather than advising - spending money, purchasing compute, signing agreements, resolving disputes - the absence of a verifiable record stops being a compliance nuisance and becomes a fundamental business risk. An autonomous transaction without a receipt is an unenforceable transaction. Somebody has to issue the receipts.
The clearest current evidence that this is the live question, rather than my opinion about it, is where the market is repricing. In the last week of July 2026, more than $1 trillion in market capitalization came off AI chip and memory names - Nvidia, SK Hynix, Samsung, Micron, AMD, ASML, Arm, SoftBank - despite several of those companies posting record profits. Infrastructure announcements no longer automatically lift semiconductor shares while investors remain uncertain about financing costs, utilization, and the revenue those assets eventually generate.
That is the market pricing raw compute as a commodity whose returns must now be demonstrated rather than assumed. Demonstration is the product this thesis is about.
The framework
Five layers, and the one structural rule that matters.
The landscape reads as five layers, and the layer a company occupies tells you most of what you need to know about its defensibility.
-
1
Physical compute and distributed infrastructure
The machines doing the work, whether hyperscale, distributed, or edge.
-
2
Cryptographic verification and trust
The systems proving the work happened as claimed: signatures, tamper-evident logs, zero-knowledge proofs, attestation rooted in hardware.
-
3
Agentic coordination and protocol
The rules by which agents reach tools and each other. This layer is being locked in right now, which is why protocol-native architecture is a screening criterion rather than a nice-to-have.
-
4
Vertical AI agents and application
Domain-deep software with data or workflow access that generalist models do not have.
-
5
Edge compute, behavioral data, and embodied AI
Where AI meets the physical world, and where persistent models of real behavior accumulate.
The companies worth backing rarely live in one layer. They sit at the intersection of two or more, or they occupy a position everything else must pass through.
Where it stands - July 2026
One leg stronger, two weaker, and a new bear case.
This is the section that makes the thesis worth reading, because it is the section most theses do not have.
Stronger · Verification
Serious institutions are formalizing the primitive. The FIDO Alliance chartered an Agentic Authentication working group with Google, Mastercard, Visa, CVS Health, OpenAI, Amazon, and Okta in leadership positions. Mastercard shipped a Verifiable Intent standard. Experian shipped Agent Trust. An IETF working-group-forming session on agent communication protocols met in Vienna in July.
When Visa, Mastercard, Google, and the IETF independently start building the same primitive in the same quarter, the category is real.
Weaker · Regulation
The forcing function just slipped sixteen months. The EU's Digital Omnibus on AI entered into force on July 27, 2026 and defers the AI Act's high-risk obligations from August 2026 to December 2027.
Any business case premised on enterprises buying audit infrastructure to meet an August 2026 deadline has to be rebuilt. The logging requirements were postponed, not weakened - and the US picture is worse, with no federal explainability mandate and several state rules actively receding.
Weaker · Crypto rails
The crypto-native leg is the weakest part of the evidence base. Of 338 tracked AI-crypto projects, 70% are already dead - measured, not projected - with average survival of 140 days. The DePIN sector has contracted sharply from cycle highs.
My revision: verification and accountability should be decoupled from crypto-native infrastructure. Token incentives must now be argued as necessary rather than assumed as natural. In most cases I no longer think they are.
New bear case · Hyperscalers
The platforms are absorbing the trust layer. Google Cloud now ships cryptographically signed agent cards, zero-trust architecture across decentralized agent systems, and IAM-integrated audit logging as native platform features. Not add-ons. Included.
This is the strongest argument against my own thesis, and it deserves a direct answer rather than a deflection. Mine is below.
My answer to the bear case
What survives above platform-native logging is the verification a platform cannot credibly perform on itself: cross-vendor attestation, adversarial or third-party audit, physical-world provenance originating outside any cloud, and records built to be admissible to a regulator or a court rather than merely useful to an engineer.
A hyperscaler attesting to its own compute is a self-certification. Some buyers will accept that. Regulated and adversarial contexts will not. That distinction - self-attestation versus independent attestation - is now the sharpest question I ask of any company in this space.
A correction to the consensus
The pricing revolution is about half as far along as it's described.
A genuine shift is underway from selling software access to selling completed work - a resolved support ticket, a drafted document, a written clinical note. That part is real.
But I went and checked the four companies most often cited as proof. Two of the four flagship examples of outcome-based pricing are conventional per-seat SaaS. One charges roughly $1,200 per lawyer per month on an annual commitment with seat minimums. Another charges about $2,500 per clinician per year. At a third, the most popular plan bills per conversation regardless of whether anything gets resolved.
The businesses are excellent and growing fast. But Gartner's much-quoted 40% figure is a 2030 projection, not a current state, and it says "at least 40%," which makes it a floor rather than a forecast.
The consequence is direct, and it is the kind of thing I get paid to say out loud: when a company claims outcome-based pricing, check the invoice. A vendor genuinely staking its margin on results is making a far stronger statement about its own confidence than one charging per seat while using outcome language in its deck.
Being straight with you
What's on this page, and what isn't.
This page gives you the framework - what I look for and why, and where the evidence currently sits. It deliberately does not give you the screen.
The screen is the part that turns a point of view into a decision: the disqualifying signals, the thresholds, the weighting by layer position, the diligence question set, and the application of all of it to actual companies. That's the work, and it's what clients pay for.
I'd rather tell you that plainly than pretend this page is the whole thing. The framework is the marketing. The screen is the product.
Work with me
Start with a conversation. That one's free.
Bring me the thing you're actually worried about - not the pitch, the part you'd skip if an investor asked. Thirty minutes, no charge, and a straight answer at the end of it.
I don't do a discovery call that's secretly a sales call. If I can be useful in half an hour, that's a good outcome and you owe me nothing. If it turns out I'm the wrong person for your problem, I'll say so and point you at whoever isn't.
If you want the whole thing
The thesis engagement
The free call is about your company. This is about the market - the full framework, the screen I actually run, and where I think value is accumulating. It's the right thing to buy if you're deploying capital, positioning a company inside this landscape, or you simply want the reasoning behind the board rather than the summary of it.
$500One hour · Includes the full written thesis
An hour on your company, your market, or your portfolio - whichever is the thing keeping you up. And the complete Trust Stack report as it stands on the day you pay, not a version I wrote in the spring and forgot to check.
- The full written thesis, including the material that isn't on this page - the screening framework in its operational form, with the thresholds and the disqualifying signals.
- An opportunity-area walkthrough covering each layer of the stack, with illustrative profiles of companies I've screened - anonymized, and offered as worked examples of what passes and what fails, not as recommendations to invest.
- The metric integrity log - every figure, its primary source, its date, and the corrections I've had to make to my own prior work. Including the one I got wrong by a factor of about 170.
- The hour itself, which is the part that actually changes anything. The report is the artifact; the conversation is the product.
Pick a time and pay in one step. Credit or debit card, PayPal, or Venmo. Your booking is confirmed once payment clears.
No hurry on that second one. Take the free call first, or read this week's signal board - it's free, it updates every Thursday, and it'll tell you fairly quickly whether I think in a way that's useful to you.