Bitcoin Brief

Sup, freaks.

The most interesting AI lab in the world right now may be the one American policymakers tried to starve of chips.

DeepSeek is not acting like a company desperate to own every interface, every customer, and every layer of the stack. It is focused on language models, agents, efficiency, and open weights. Its founder sounds calm. The American frontier labs sound increasingly frantic.

That contrast matters more than another benchmark leaderboard.

Let's get into it.

LEAD STORY

DeepSeek Is Playing the Long Game

Leaked, audio-derived transcripts of a May 20 investor meeting with DeepSeek founder Liang Wenfeng have been circulating this week. The recording runs nearly four hours. DeepSeek has not publicly authenticated it, and the available English texts rely on automated transcription and AI-assisted translation. Yicai sought confirmation and reported that the company had not responded. That means the exact wording deserves caution.

The strategic posture does not.

Liang describes AGI as DeepSeek's long-term goal and the organizing priority behind its research decisions. Not winning the consumer traffic war. Not becoming the operating system for every enterprise. Not vertically integrating into every downstream market. Not bolting on a video generator because investors are excited about video this quarter.

Build the best language model possible. Push it into agents. Solve continual learning so those agents can retain knowledge instead of waking up as amnesiac contractors every morning. Liang said DeepSeek would probably continue open-sourcing even its strongest models. He described API pricing that targets recovery of equipment purchase costs in roughly ten months rather than maximum revenue. And he said he hopes partners, not DeepSeek, will build most downstream applications.

It is an almost offensively restrained strategy.

Then look west.

Dario Amodei and Sam Altman talk constantly about missions, safety, and serving humanity. From the outside looking in, Anthropic and OpenAI increasingly look like they are racing for control. They want the model, the interface, the enterprise relationship, the coding agent, the consumer account, the distribution, and the policy moat. They invoke ideology when it helps build trust, then pursue enclosure when it helps consolidate power.

DeepSeek may ultimately behave the same way. No founder transcript is a binding constitution. The company operates in China, under a political system that hardly deserves romantic treatment. Open weights are not the same thing as open data, open training code, or political freedom.

But on the evidence in front of us, Liang sounds less frantic than the American frontier CEOs. He sounds like a founder with a long time horizon who knows exactly which layer he wants to own and which layers he is willing to leave alone.

I am rooting for that mindset.

There is an awkward US-versus-China tension in saying so. Fine. If a Chinese lab is demonstrating more restraint, more patience, and more commitment to open models than America's supposed champions, the American response should be to learn from it, not hide behind nationalism.

The chip constraint makes the comparison even more important.

Liang's argument is that China's primary disadvantage is access to compute, not a shortage of capable researchers. He ties DeepSeek's emphasis on efficiency partly to that constraint: when compute is limited, better computational efficiency lets the lab train larger models. Export controls and limited access to the best chips sharpened the incentive to obsess over architecture, training efficiency, inference efficiency, and deployment economics.

That can become an advantage.

At PubKey's Energy and Compute event in Washington last week, I was asked whether chips or power would become the bigger constraint. My answer was that a chip shortage is more sustainable than a power shortage. There are enormous efficiency gains still available in model architecture, training, inference, hardware, software, and implementation. A company can do more intelligence per chip. It cannot negotiate with a grid connection that does not exist.

DeepSeek is the clearest live example of that thesis. Scarcity sharpened the lab's incentive to find efficiencies that better-supplied competitors had less reason to prioritize. Architecture, culture, and cost discipline mattered too. American labs could spend their way around waste. DeepSeek had to engineer around it.

This does not mean chips stop mattering. It means the market may be underestimating how quickly useful intelligence can compound on constrained hardware. It also means analysts should stop treating today's GPU requirements as a fixed law of nature. The training stack, inference stack, model architecture, and application layer are all moving at once.

The deeper contrast is cultural.

One side looks determined to own the entire intelligence supply chain before the window closes. The other is acting as if the work itself matters more than winning every adjacent market today.

To me, that reads as low time preference.

Bitcoiners should recognize it immediately.

SIGNAL

One Standard for Iran. Another for the Client.

On July 22, Energy Secretary Chris Wright and Saudi Energy Minister Prince Abdulaziz bin Salman signed a US-Saudi civil nuclear package: a Section 123 agreement plus bilateral safeguards. DOE sold it as a multi-billion-dollar nonproliferation-friendly partnership.

The fuel cycle is the fight. ABC News, citing two US officials, said the deal leaves open future domestic Saudi enrichment and described safeguards as bilaterally run rather than classic IAEA-led oversight, far from the UAE model that forswore enrichment.

Then the cleanup: Trump said there would be "no enrichment of material" and made the deal contingent on Saudi entry into the Abraham Accords. The White House backed that contingency while admitting he had not yet raised it with Mohammed bin Salman. Signed framework, enrichment-pathway reporting, presidential denial, late condition, Congress still ahead. Iranian enrichment is destiny. Client rules flex. Language is easy. Discipline is expensive.

Jensen Just Normalized Distillation

Rohan Paul circulated a clean Jensen Huang clip on the question closed labs love to moralize: should open models distill closed ones?

Jensen refused the small-frame answer. Learning from others is foundational to intelligence. Humans do it constantly. Early models trained on human internet data. Future systems will train mostly on other AI systems as the corpus itself becomes machine-written. Distillation is not a weird edge case. It is how intelligence compounds once models become the authors.

If frontier labs want permanent rents on cognition, they need law, distribution control, and API terms to do work that learning will not do for them. Jensen also inverted the safety script: smarter systems, in his framing, are likelier to be safer. That will not end the alignment fight. It does puncture the idea that freezing learning pathways is automatically the responsible move.

Bottom-Heavy, Not Bullish Yet

James Check is not begging you to smash buy. He is pointing to an ownership transfer under the surface.

Bitcoin recovered the 200-week moving average around $63k and held it for four consecutive weeks. Check estimates roughly $100 billion changed hands between $58k and the mid-$60s while long-term holders accumulated more than 80% of realised capital. ETFs and Strategy were sellers during part of that defense, meaning less visible buyers absorbed the coins.

The divergences are constructive. Weekly RSI made a higher low. Short-term holder MVRV and realised-loss intensity are improving. But Check's rule still stands: do not remove the bear goggles until the short-term holder cost basis near $68k is meaningfully cleared and held.

The bull case is not euphoria. It is absorption.

They're Not Storming Crimea. They're Switching It Off.

Caolan Robertson's interview with Ukrainian unmanned-systems commander Robert Brovdi (Madyar) is doctrine, not theater.

Occupied Crimea is not self-sufficient. Brovdi wants occupation to fail by killing inputs: power, fuel, logistics. Leave exits open. Make the military machine unable to stay. The same logic reaches into Russia proper. Cancelled flights and fuel stress are the point. Cheap long-range drones collapse the bargain that Ukraine burns while Russian daily life stays normal.

The industrial root is simple. A civilian-heavy unit near Kherson in 2022 could not see the artillery hitting it. Commercial drones became recon, then armed craft, then FPV, then a national unmanned force. Brovdi's NATO homework: border drone belts, fiber-controlled mass launchers, cheap counter-drone defense, real unmanned brigades, and deep-strike crews now. Exquisite interceptors lose the math. The side that industrializes first rewrites the map.

The ETF Bid Just Blinked

After several green sessions, US spot bitcoin ETFs printed a $225.2 million net outflow on July 23. Total net assets sit around $78.8 billion, with $51.6 billion in cumulative net inflows since January 2024.

This does not kill Check's ownership thesis. It sharpens it.

Public tape can wobble while private absorption continues. It also means nobody gets to pretend the market has entered a one-way ETF melt-up. The bid is real on some days and absent on others. Bottoming processes look like this when leverage is lower and conviction capital matters more than slogan capital.

⚡ FREEDOM TECH CORNER

If It Works During a Shutdown, the State Calls It a Threat

Jack Dorsey said the quiet part out loud: the government of India does not like technologies like bitchat and wants it taken down.

Bitchat is offline-first messaging over Bluetooth mesh. No internet required. No phone number. No central server. Devices relay encrypted messages across physical proximity networks. That architecture is annoying to product managers and intolerable to shutdown regimes.

India has used internet blackouts as crowd-control infrastructure for years. A messenger that keeps working when the ISP kill switch flips is not a novelty app in that context. It is a countermeasure.

Governments will always have a public-order story ready for tools like this. Anarchists. Rumors. Investigation interference. The pattern is older than smartphones. Communication that the state cannot mute gets redefined as dangerous the moment it works at the wrong time.

The important distinction is hard power versus theater. Pressuring GitHub or app stores can slow distribution. It does not erase the category. Once people learn that mesh and offline relay are possible, the policy demand becomes permanent.

If your network only exists by permission, it is not a network.

It is a lease.

DATA SNAPSHOT

As of July 24, 2026, morning ET

See you tomorrow. This is not investment advice. Do your own research.

YouTube: https://www.youtube.com/@TFTC

podcast: https://www.tftc.io/tag/podcasts/