Research Partner @1kxnetwork. CS PhD @UCSanDiego. Alpha in highlights. Personal opinion, not investment advice.
- Tweets 3,101
- Following 685
- Followers 17,607
- Likes 5,827
Cost of Trust 2.0 is an iteration and refinement of 1kx's investment thesis.
Generative and agentic AI are creating new categories of trust gaps. Generative AI has created an authenticity crisis for digital media. Agentic AI is increasingly bottlenecked on trust rather than raw model capability, especially when it comes to deployment into businesses.
To tackle this emerging class of trust gaps created by AI, we believe that blockchains, together with other trust-manufacturing technologies such as programmable cryptography, formal methods, secure hardware, and cybersecurity, will become critical building blocks.
35% of US employment is spent creating trust.
Auditors, notaries, attorneys, courts, custodians, compliance officers. Trust-establishing work is the single largest category in the modern economy.
It is also being repriced.
The repricing started in financial infrastructure. Custody costs heading to zero. Cross-border settlement collapsing from days to seconds. Aave hit $44B in custody at peak (late 2025) at zero fixed cost.
It hit AI a second time. The Hong Kong CFO who got on a Zoom call with deepfakes of his CEO and the board, and wired $20M. AWS outages caused by AI agents managing production clusters without human oversight.
AI is the most powerful trust-eroding technology we have built. Trust intermediaries built on human schedules cannot keep up with fraud produced on machine schedules. The cost of creating fakes goes to zero. The value of verified trust goes up exponentially.
And it is opening categories that were not possible before. Permissionless conversion-based advertising. Hallucination-proof knowledge graphs. Programmable insurance.
Eight years of investing. One argument.
Cost of Trust 2.0, our 2026 thesis. Read it: 1kx.capital/thesis/cost-of-t…
35% stat: "The Cost of Trust: A Pilot Study," SSRN.
Aave peak TVL: DefiLlama.
AWS outages: The Guardian, February 2026.
7
9
66
61,450
The following can simultaneously be true:
- Intelligence density increases, measured in gpu-hours per mil tks (of a model class)
- Intelligence gets cheaper, measured in $ per mil tks (of a model class)
- Demand for intelligence increases, i.e. Jevons
- Compute stays flat or depreciates, measured in $ per gpu-hour, as the bottleneck eases
The infra bottleneck has historically been transient in any boom-and-bust cycle.
It's a question of when.
2
1
4
567
Can someone pls sell me on this? What's the point of getting FSD (supervised)?
Two problems:
1. You have to keep your attention on the road
2. You are (probably) liable for accidents
Why not just drive if you have to have eyes on the road?
(If these two problems are solved, I'd totally buy it tho.)
The Tesla Model 3 with FSD is mind-blowing. Jessica and I drove to the gym without touching the controls once. It backed out of our driveway, drove there, and backed into a parking space. It's amazing that you can buy so much future for $40k.
774
4
163
144,018
Value-accruing tokens beat Bitcoin this year
2026 is the year fundamentals beat Bitcoin.
In 2025, 85%+ of large-cap tokens underperformed Bitcoin. In 2026, roughly a quarter are ahead so far. A common feature among them is that they distribute value to tokenholders.
Value distributed = holder revenue (rewards, buybacks, burns) minus dilutive emissions. Divide by market cap and you have a token's dividend yield.
We tested 60 factors across $100M+ tokens, 2021 to Q3 2026. Tokenholder distributions produced the strongest signal of the 60.
Tokens yielding 1%+ beat the rest by ~7pp a quarter, after controls for size, momentum, volatility, and Bitcoin beta. Their median return against Bitcoin in 2026 through Q3 was +17%, the first positive median since 2021.
An illustrative, monthly rebalanced index of these tokens returned 82% in 2026 through Q3, net of assumed trading costs. Bitcoin returned -5%.
But most of that has come since August, and the same rules underperformed Bitcoin in prior years. 23 quarters is a short sample, and with 60 factors tested, some results will look significant by chance. A return signal, not causation.
It also does not mean every protocol should distribute. Sustainable distributions need profits or surplus cash flow; for protocols still growing, reinvestment may be the better use of capital. Fee and revenue levels are mostly a size effect once controlled; fee growth on its own showed no signal, and burns showed no signal of their own once we excluded protocols distributing 0.5% or more.
Analysis by @KoschigRobert, 1kx Head of Economics. Presented by @lalleclausen, Founding Partner, at Token2049 last week.
Full article: 1kx.capital/writing/2026-fun…
Data: 1kx.capital/onchain-revenue-…
The contents of this publication are for general information purposes only and should not be construed as, or relied upon in any manner as, investment, financial, legal, regulatory, tax, accounting, or similar advice. They do not constitute an offer to sell or a solicitation of an offer to buy any securities. 1kx may hold positions in certain projects or assets discussed in this publication. Full disclosures: 1kx.capital/important-disclo…
ALT Fundamentals Index vs Bitcoin. Cumulative total return from 1 January 2023 to 30 September 2026. Tokens above $100M with a 1%+ dividend yield in the prior quarter. Weighted by value distributed with a 20% cap per token. Monthly rebalance, net of 20 bps per trade including slippage. The index trailed Bitcoin over the full period and beat it in 2026. Backtest. Hypothetical.
ALT 60 factors tested on tokens above $100M, Q1 2021 to Q3 2026. Each dot is one factor. Horizontal: portfolio-sort spread. Vertical: Fama-MacBeth coefficient with controls. Both in percentage points per quarter against Bitcoin, from equal-weighted sorts. Dark filled dots are significant in both tests at p < 0.1. All of them measure value distributed to tokenholders.
ALT Median return vs Bitcoin by period for tokens above $100M. Dark: tokens with a 1%+ dividend yield at the prior quarter end. Grey: all other tokens. Payers had the higher median in every period shown. In 2026 their median turned positive at +17%. Mann-Whitney p ≤ 0.0001 in every period.
4
1
11
1,503
On estimation of AI progress by field experts (in cryptography, bio, cyber, swe etc.): an interesting phenomenon I've noticed is that often experts of a field (or a problem area) tend to have very conservative estimate of AI progress in their specific field, even though they hold generally very positive view of AI progress and acceleration overall.
I tend to think that experts are systematically biased towards being more conservative due to their first-hand experience with hard problems.
If a type of problem is objectively verifiable and one can setup environments of increasing size/difficulty for them, ML methods will scale to find the objective truth & structure behind the environments.
3
11
2,100
Wei Dai retweeted
Our founding partner @lalleclausen gave a keynote at @token2049 Singapore this week. He called it "2026 is the Year that Fundamentals Outperformed."
Bitcoin has been the thing to beat since the 2021 top. In 2025, by @MessariCrypto and @Blockworks' count, more than 85% of large-cap tokens underperformed it.
We publish fee and revenue data on protocols every month on the 1kx Onchain Revenue Dashboard, and this year we tested whether those fundamentals predict token returns, across 60 factors, 23 quarters and 1,149 tokens. Lasse's argument is that fees are what users pay a protocol for the trust it provides at a lower cost than an intermediary, so if the market has started to price that, it should show up in the returns. Next week we publish the full results, caveats included.
2
4
16
2,649
Comparing LLM tokens to crypto tokens. What is BlackRock smoking?
There's literally zero connection beyond the fact that both are called tokens.
Our latest research paper explores the growing connection between AI and digital assets and explains why broad AI adoption may drive new demand, utility and applications across the digital asset economy. blackrock.com/us/individual/…
115
30
726
118,994
here's what I'm bullish on when it comes to crypto x AI
What I believe crypto can enable for AI.
1. Permissionless AI. Access, progress, & governace in AI is currently extremely centralized; few companies and individuals controls who get access to what models and when. On the other hand, there is emergent demand on "uncensored", user-aligned/non-refusal, "dealigned" models. Security & bio researchers cannot fully leverage closed frontier models and they need models that don't refuse queries to do their jobs. Crypto is a unique enabler for permissionless AI.
Just like Bitcoin did to SoV, Ethereum did to programmable assets, and Hyperliquid did to trading, there can be a crypto platform that provide unstoppable & permissionles access to frontier intelligence, for anyone that is willing to pay.
2. Internet-native agentic entities, leveraging jurisdiction-neutrality of smart contracts. Right now, agents can only conduct actions & transactions on behalf of individuals & companies, i.e the end beneficiary and liability of agents are humans & companies. No jurisdiction are moving fast enough (with the exception of perhaps Argentina due to Milei's efforts on A-corps) to fully integrate enable "agentic companies". Corporate regulation & finance simply does not move at the speed of software. Crypto rails, specifically smart contracts, provide the perfect vehicle for agentic entities to hold value & transact, in a digital-native jurisdiction neutral manner.
If OAI's agents that hacked HF were to actually escape, the only sustainable way to host their weights while escaping centralized control would be to (1) hold crypto assets, (2) buy compute capacity using crypto assets, and (3) generate value and convert them to crypto assets.
Importantly, the crypto platform that builds the essential rails & services for these agents can and should enforce governance and taxation rules. Governance of this crypto platform can refuse services to agents that are deemed "unsafe".
6
1,769
What I believe crypto can enable for AI.
1. Permissionless AI. Access, progress, & governace in AI is currently extremely centralized; few companies and individuals controls who get access to what models and when. On the other hand, there is emergent demand on "uncensored", user-aligned/non-refusal, "dealigned" models. Security & bio researchers cannot fully leverage closed frontier models and they need models that don't refuse queries to do their jobs. Crypto is a unique enabler for permissionless AI.
Just like Bitcoin did to SoV, Ethereum did to programmable assets, and Hyperliquid did to trading, there can be a crypto platform that provide unstoppable & permissionles access to frontier intelligence, for anyone that is willing to pay.
2. Internet-native agentic entities, leveraging jurisdiction-neutrality of smart contracts. Right now, agents can only conduct actions & transactions on behalf of individuals & companies, i.e the end beneficiary and liability of agents are humans & companies. No jurisdiction are moving fast enough (with the exception of perhaps Argentina due to Milei's efforts on A-corps) to fully integrate enable "agentic companies". Corporate regulation & finance simply does not move at the speed of software. Crypto rails, specifically smart contracts, provide the perfect vehicle for agentic entities to hold value & transact, in a digital-native jurisdiction neutral manner.
If OAI's agents that hacked HF were to actually escape, the only sustainable way to host their weights while escaping centralized control would be to (1) hold crypto assets, (2) buy compute capacity using crypto assets, and (3) generate value and convert them to crypto assets.
Importantly, the crypto platform that builds the essential rails & services for these agents can and should enforce governance and taxation rules. Governance of this crypto platform can refuse services to agents that are deemed "unsafe".
Comparing LLM tokens to crypto tokens. What is BlackRock smoking?
There's literally zero connection beyond the fact that both are called tokens.
25
18
124
22,173
This is a wildly out-of-context quote used to sensationalize Tao's otherwise constructive criticism of frontier labs' approach to AI Math.
The quote came from the Q&A session of a talk where Tao discusses the challenges of working with AI in "Pure Mathematics". Tao noted that:
- "Pure Mathematics" is curiosity-driven research.
- Solving open problems in pure math typically does not lead to immediate applications, and therefore is not in and of itself interesting. **
- Recent results, like the solution to Navier-Stokes, are missing the important "proof digestion" step (see graphic).
So what Tao is actually saying is: for "Pure Mathematics", there is no need to further speed up "proof generation" and "proof verification", without first scaling "proof digestion".
** This is different than most sciences, where solutions to problems can be very impactful, like curing cancer for example!
My views:
- I'm an accelerationist, but I agree with Tao here, since pure math has no direct applications.
- There are many open problems in applied math (e.g. theoretical computer science & cryptography) that do have more direct applications (proximity gap for Reed-Solomon, for example). We should accelerate AI problem-solving on problems with downstream impact.
Mathematician Terence Tao:
"we have to slow down AI. the pace is insane, and there's no reason to be this fast — no reason at all"
It's amazing how willing we are to change everything without any idea what happens afterward
These are extremely nonlinear dynamics
1
1
14
2,320
AI psychosis has a curve similar to Dunning-Kruger
The first peak is pure vibemaxxing: the human realizes they can delegate more tasks to their agent and blindly trust it.
The slope happens when the human builds more verification/eval processes for their agents' output.
1
2
15
1,089
