TFTC

Bitcoin Brief

Sup, freaks.

SemiAnalysis just published a piece that reframes the entire AI debate. They call it Dark Output: real economic value that AI is producing right now that GDP literally cannot see. A basic legal document that cost $400 thirty years ago now costs $0.50 through AI. GDP does not register that as a productivity gain. It registers it as a decline. The incoming Fed Chairman, speaking about economic data broadly in December 2025, admitted the data is backward-looking. He was not talking about AI specifically. But the point lands harder when you apply it to Dark Output: if the Fed is already behind on conventional data, the invisible AI output makes the gap exponentially worse. The people calling AI a bubble are measuring the costs while the output is invisible to their instruments. We break down why this matters and what it means for Bitcoin.

Lead Story

Most of What AI Produces Is Invisible to the Economy That Paid for It.

SemiAnalysis (Dylan Patel, Malcolm Spittler) published AI Dark Output on May 29. Core thesis: AI is creating real economic output that national accounts cannot see. Like dark energy in physics, it will only be visible through its effects on other things. The framework explains why the AI debate feels so broken. The bulls see real productivity gains. The bears see massive spending with unclear returns. They are both right and measuring different sides of the same phenomenon.

Three types of Dark Output. First, substitution: AI does work humans used to do, receipts vanish from GDP. A will went from $400 to $150 over 30 years (GDP handled it). $150 to $0.50 in one year vanishes from the dataset entirely. Second, new work that was too expensive until AI made it cheap. Literature reviews, meeting dossiers, email summaries. Creates real value but no economic trace beyond API spend. Third, captured work where companies have market power and charge the same price for AI-produced work. Margins explode but the work is identical.

The fingerprint they found: employment in AI-exposed sectors falling relative to broader economy while wages in those same sectors RISE. Not because anyone got a raise. Because the cheapest workers left the sample. Junior staff displaced first, average wage moves up mechanically. Fed sees rising wages, thinks labor market is tight. In reality AI ate the bottom of the labor stack. The next Fed chair admitted it. Incoming Fed Chairman Kevin Warsh (December 2025): "If you are looking at the data, my view is you are backward looking. You are going to be late." He is telling you the instruments are broken.

We are a case study of Dark Output. In today's TFTC episode with Jordi Visser, I described the company brain architecture: a team of five or six people running OpenClaw since January with a persistent memory system that gives everyone access to every transcript, every newsletter, every financial record, every ad partner interaction. Visser's reaction: "This is an incredible validation of what you're building." He pointed out that Torsten Slock is making the argument most economists reject: the entire US economy right now is AI. No job creation, but consumption growing anyway. The AI side is allowing profit margins to grow while the labor data shows nothing. That is Dark Output in Slock's framework too. A team of less than 10 people producing output that would require a team roughly double the size without AI. None of that productivity shows up in any macro statistic. The Brief, the podcast, the research, the interview prep. All of it is new Dark Output: work that was too expensive to do until AI made it cheap. Multiply that across every company implementing agents and you get the measurement gap SemiAnalysis is describing. The economy is more productive than it appears, but that productivity is invisible to the systems that measure it.

The deeper question nobody has answered yet: how do you actually measure the productivity growth? I can think of dozens of tasks people are running through these tools right now that would never have been done before but are being done. How do you turn that into economic value? How do you express it in a number that connects to the work? It is genuinely hard to tell.

We will figure it out over time, but in the interim, a lot of bad calls are going to get made because people are looking at old measurements for a new economy. This is the core of Visser's thesis that he has been developing over the past year: you have to throw out the old economic models and measurement tools because things are changing faster than the instruments can track.

The question is how companies cross the implementation chasm. As Visser put it, asking "how do I start with AI" is like asking how to lose 30 pounds. Everyone knows the answer. Almost nobody does it. The companies that do, the ones producing Dark Output right now, have an advantage that widens every day the measurement gap persists. The macro instruments cannot see them winning. By the time the data catches up, the game is already decided.

For Bitcoin: if AI output is real but invisible, three things follow. First, the people calling AI a bubble are measuring the wrong side of the ledger. Second, the Fed is making decisions with broken instruments. Third, the tax base that governments depend on (wages, transactions, receipts) is shrinking while actual economic activity is not. That gap has implications for the case for holding assets that exist outside the measurement system entirely.

Signal

Private Credit

Cliffwater's $33 Billion Fund: 17% Want Out. 5% Are Allowed to Leave.

Why it matters: Q2 redemption requests nearly tripled. The fund responded by tightening the gate. They'll honor less than a third of what investors want out.

The private credit crack is starting to show. Cliffwater Direct Lending, managing $33 billion and the largest independent name in the space, saw redemption requests double in Q2. Their response was to cut the quarterly gate from 7% to 5%. Translation: if 17% of investors want out, only 5% will be allowed to leave. The rest get queued. This is the same playbook real estate funds used in 2022. Gate down, hope the pressure subsides. It rarely does.

Bitcoin Protocol

Peter Todd: Why Zcash-Style Privacy at Bitcoin's Consensus Layer Is a Bad Idea.

Why it matters: Zcash just had an emergency fork this week that proves the point. Todd argues Zcash-style cryptography carries fundamentally higher risk than Bitcoin's current primitives. This week proved him right.

Todd's argument: zero-knowledge proofs at Bitcoin's consensus layer are a net negative. Higher complexity, more attack surface, and the failure mode is catastrophic. Privacy is valuable, but base layer security is more valuable. Privacy can be built on layers above. Security cannot be added after the fact.

During routine auditing, an issue affecting Zcash's Orchard shielded pool was identified over the weekend. ZODL's disclosure is worth reading carefully because the language tells you everything: Orchard-related transactions stopped being mined. Privacy was unaffected. All funds were safe. The issue was caught before any known exploitation. But it required a protocol-level change, which took effect at 22:30 EDT on June 1.

Only Orchard, Zcash's latest and most advanced shielded pool, was affected. Sapling and transparent transactions continued. ZEC on exchanges remained tradable. But wallet users could not send or receive Orchard funds until the upgrade was complete. ZODL also notified maintainers of other protocols that have deployed Orchard as part of responsible disclosure, meaning the vulnerability surface extends beyond Zcash itself.

The bug was a soundness vulnerability in the Orchard zero-knowledge proof circuit, specifically in the halo2_gadgets crate. "Soundness" means the protocol should only accept valid transactions and valid state changes. A soundness bug allows the system to accept something it should reject. Successful exploitation could have allowed invalid state transitions inside Orchard, potentially breaking the accounting guarantees for the entire pool.

The fix required two stages: a soft fork to disable Orchard entirely (because a direct patch would have revealed too much about the vulnerability), followed by a hard fork to update the ZK proof circuit itself. The whole process relied on "voluntary cooperation among independent participants throughout the network." This is only the second security-driven protocol upgrade in Zcash history since launch in 2016. Bitcoin's conservative approach to cryptographic upgrades exists precisely to avoid this failure mode.

16 Headwinds for AI. Revenue Is Slowing. 0DTE Options Create 1987-Style Systemic Risk.

Why it matters: Brandon Carl's thread documenting the bear case. The bear case for AI infrastructure, from someone who was bullish before. The counterpoint to the Dark Output bull case.

Brandon Carl documented the AI bear case in 16 points this week, and it deserves a full read. Carl is not a perma-bear. He scaled AI to tens of billions of communications and his teams were experimenting with attention mechanisms back in 2017. When he turns cautious, it carries weight.

The core argument: input costs are higher (commodities, chips, power), interest rates are higher, and competition is more intense than when the original investment theses were written. Revenue growth appears to be slowing. The technology is valuable, but ROI is proving more expensive and taking longer than anticipated. Scaling laws are becoming problematic: exponential costs and power consumption cannot continue on the current trajectory.

The financial risk angle is where it gets pointed. Leverage is substantially higher than in previous cycles: leveraged ETFs, call option activity, and margin loans are all at extremes. Korea is particularly exposed. Zero-day-to-expiry options create a profile with stronger parallels to portfolio insurance and the 1987 crash than any other point Carl can remember. The combination of exponential increases in call activity coupled with the ties of semiconductors to structured products means there is, in his words, a non-trivial systemic risk to the financial system.

On the technology side: current memory maker forecasts are built on quadratic attention, but DeepSeek, Minimax, and NVIDIA are already showing work that can cut RAM needs by 80% or more. That means semiconductor valuations are pricing in demand that algorithmic improvements may eliminate. The future is likely "different models for different use cases" with the lower end of the market being highly competitive. The bull case is Dark Output. The bear case is that the math does not work at current valuations. Both can be true.

Markets

Checkonchain: Red Waterfall. Net Realized Losses Hit $2.06 Billion.

Why it matters: Fourth Checkonchain piece this week. The sell-off is accelerating.

James Check called it Red Waterfall. This is his fourth piece this week, after Peak Apathy, Stay The Course, and Slaying The Sacred Cow. The pace of his publishing tells you what the data looks like.

The on-chain picture deteriorated sharply overnight. Net realized losses hit $2.06 billion, the worst reading since the February capitulation. LTH-SOPR dropped to 0.754, meaning long-term holders who have been through multiple cycles are now spending coins at a 25% loss on average. That is deep capitulation territory. MVRV Z-Score fell to 0.43, approaching the levels that historically mark cycle bottoms.

Short-term holders are $9,400 underwater on average, with the STH realized price at $76,473 against a spot price of $67,000. NUPL has fallen to 0.187, entering the fear zone. In the February episode with Marty, James laid out his framework: if you believe Bitcoin is dead, this has no floor. If you believe it is not dead, we are in the bottom 20% of every mean reversion model. The data today says we may be deeper than that. His bison analogy from that conversation applies: turn and face the storm. It is the fastest way to get to the other side.

The question James keeps asking: are we seeing capitulation or distribution? Capitulation ends when the sellers are exhausted. Distribution means there are more to come. The answer usually shows up in whether long-term holders in profit start selling alongside those in loss. Watch that signal in the coming days.

Commodities

HSBC Warns of 'Super-Squeeze.' Goldman Hikes Copper to $13,735.

Why it matters: Goldman raised year-end copper forecasts. HSBC warns of a commodity super-squeeze.

HSBC warned of a commodity "super-squeeze" this week. Goldman Sachs raised copper forecasts to $13,735 per ton. Mine supply forecasts were cut by 350,000 tons. US copper imports are running above forecasts as companies front-run potential tariffs. Hormuz remains closed, affecting energy costs for mining operations. Goldman mapped three scenarios: base case $11,000, upside $13,735, extreme case (if supply disruptions worsen) $16,000. Copper is the backbone of the energy transition and AI infrastructure buildout. Both require massive amounts of it.

Two Small AI Models Just Did What Neither Could Do Alone.

Why it matters: Qwen 35B as the brain + LocateAnything 3B as the eyes filled out a paper form together. Capability through composition, not scale. The compute cost curve bends further.

Steve Kovach demonstrated something important this week. Qwen 35B (reasoning) + LocateAnything 3B (vision) working together filled out a complex paper form. Neither model could complete the task alone. The combination created capability that exceeded the sum of the parts. This is capability through composition, not scale. If you can get frontier performance by combining small, cheap models instead of training one massive expensive model, the entire cost structure of AI shifts. The compute requirements drop. The accessibility increases.

⚡ Freedom Tech Corner

Run Qwen 3.6 Locally. No Cloud. No API. No Surveillance.

The small model story in today's signal section isn't theoretical. You can run Qwen 3.6 on your own hardware today. Download it from Hugging Face, run it with llama.cpp or Ollama, and you have a capable AI that never sends your data to anyone. No API key. No subscription. No cloud provider reading your prompts. This is what sovereignty looks like in the AI era.

If you have a machine with 16GB of RAM, you can run the 8B parameter version. If you have 32GB, run the 35B. Start here: https://huggingface.co/Qwen

Data Snapshot

Bitcoin Price	$67,000
Sats per Dollar	1,493
Block Height	952,202
Hashrate	759.8 EH/s
Daily Fees	$273,256

MVRV Z-Score	0.43 Approaching undervalued
LTH-SOPR	0.754 Deep capitulation
STH-SOPR	0.972 STH spending at a loss
STH Realized Price	$76,473 STH $9,400 underwater
NUPL	0.187 Fear zone
Net Realized P/L	-$2.06B Worst since February
Realized Cap	$1.08T