A Bitcoin Miner Switches to AI for a $4.7 Billion Lease

In August 2026, which layer of the AI stack would you put a dollar into today, and which layer would you avoid?

LPS (Land Power Shell) is still the most obvious and fastest path to cash-on-cash returns.

My partner and I have assembled almost 6GW of grid and behind-the-meter power coming online in a ramp from today through 2029. Energized land with a signed interconnection is the one input here that cannot be manufactured on demand, and as data centers meet more local resistance, the value of a site that is already permitted and powered goes up.

Silicon is where I would tell you to be careful, and I say that as someone who helped get Groq off the ground in 2015 until it got licensed to Nvidia for $20B last December. The capital required has grown faster than the number of defensible positions available, and you are underwriting against a competitor with unlimited resources.

The winners today and likely long term are in harnesses and the applications built on them. A harness is the software wrapped around a model that decides what the model sees, links tools, manages the execution loop, and packages workflows as repeatable skills. That is where a company can put its own data and workflows, creating a unique moat. A competitor cannot copy it by switching model providers. It matters more now than it did a year ago, because I also believe we have entered a recursive self-improvement loop. Looking at the results from the various labs over the past few weeks, I would say we are firmly in this loop now, and this means the marginal costs of all models will go to ~$0.

One caution on all of it. Right now everyone is glomming onto AI like it’s a life raft. But these same folks have not yet shown repeatable, audited, verifiable ROI even as their CapEx and OpEx are increasing with tokenmaxxing. The next 18 months will be wild, so make sure you can show your work.

What do you think?

On to the rest of the What I Read This Week…

Bitdeer spent the last several years mining bitcoin in Norway. On August 4, it signed a 16-year lease worth about $4.7B to convert its Tydal mining site into an AI data center for Volta, and the stock rose 9%.

Volta builds AI data centers and is one of Nvidia’s approved cloud partners. Its customer at Tydal is an unnamed frontier AI lab, with Dell supplying the hardware. If Volta exercises its extension options, the lease stretches from 16 to 24 years and total contracted revenue reaches ~$8B.

The tenant is Volta, which builds AI data centers as one of Nvidia’s approved cloud partners. Volta’s own customer at Tydal is a frontier AI lab that neither company has named, and Dell is supplying the hardware.

Miners spent the last decade buying cheap land near cheap power and securing agreements with utilities to connect to the grid. Those connections can take years to obtain, which is why an AI operator may prefer to lease an existing powered site rather than wait for a new one. Tydal runs on Norwegian hydropower and loses only a tenth of the electricity it draws to cooling and overhead, roughly half the waste of a typical data center.

J.P. Morgan and one other global bank are expected to guarantee about $1.3B of Volta’s payments through letters of credit, which are promises by a bank to pay the landlord if the tenant does not. That covers roughly a quarter of the contracted revenue.

The trend continues to grow as public miners have now announced tens of billions of AI and high-performance computing contracts. CoinShares expects listed miners to draw as much as 70% of their revenue from AI by the end of 2026, up from about 30% in March.

Certain AI workloads rely on a process called sampling, where computers repeatedly choose among many possible outcomes according to a probability distribution. For example, a generative model might calculate several possible outputs, assign each a probability, and then randomly select one according to those odds.

GPUs perform this sampling by running calculations that imitate randomness.

Extropic is pursuing a different approach: a Thermodynamic Sampling Unit, or TSU, that uses the physical randomness of electrical noise in ordinary transistors to make those selections directly.

The basic component is a pbit, or probabilistic bit. An ordinary digital bit is designed to remain at either 0 or 1. Extropic’s transistor-based pbit fluctuates between those states. Connecting many pbits allows a TSU to sample from programmable probability distributions.

People hear this and think of quantum computing. A quantum bit, or qubit, can exist in a quantum superposition, meaning a combination of 0 and 1, until it is measured. A pbit is always a conventional 0 or 1, but randomly switches between the two. Quantum computers go to enormous lengths to shield qubits from outside noise because it can destroy their quantum state. Extropic deliberately uses the natural electrical noise in transistors to make its pbits switch (read our Quantum Computing Primer here).

On August 3, Extropic unveiled Z1, its first scaled-up TSU. The company says the chip contains 269,568 pbits, produces more than 50 million samples per second, draws less than 1 watt, and measures under 12 millimeters on each side. Z1 has been taped out, but systems have not yet shipped to customers. Extropic plans to offer early access to Z1 cards, sticks, and clusters in 2027.

Extropic says its design can reduce energy use by processing information where it is stored instead of repeatedly moving data across a chip. Its headline claim is that it is up to 10,000x as energy-efficient as GPUs for suitable workloads. That figure is not a measurement from a deployed Z1 system, but from papers and simulations involving small denoising and generative-model benchmarks.

A week earlier, Extropic said it signed a nonbinding letter of intent with the Commerce Department for up to $75 million through the CHIPS R&D Office. If final terms are reached and milestones are met, the funding would support Z1 clusters, benchmark testing and a follow-on Z1.5 chip at a U.S. foundry.

“The current era of AI is one of exponential acceleration, and energy is its binding constraint. Brute-forcing the digital, deterministic paradigm to unfathomable scale cannot be the endgame… TSUs harness the inherent randomness of nature to deliver more intelligence per watt. And because they run on mature nodes, we can build them here, in American fabs, without waiting in line for anyone else’s capacity.” - Guillaume Verdon, founder and CEO of Extropic.

Alpamayo 2 Super is Nvidia’s open AI model for autonomous driving. It takes in what the car’s sensors see and decides what the car should do next.

Nvidia released it for free on August 4 under OpenMDW-1.1, a permissive Linux Foundation license, so any company can download it, retrain it on its own driving data, and sell the result.

Beyond steering control, Alpamayo produces a plan for its path and writes out in plain language why it chose that path, while generating labels for its own training footage. On LingoQA, a test that scores how well a model answers questions about what is happening in a driving scene, it placed first among nearly 40 systems, ahead of Gemini 2.5 Pro and GPT-4o.

The Alpamayo models have been downloaded more than 500,000 times on Hugging Face. Nvidia lists safety validation among the uses it has in mind and says the automatic labeling cuts the time required to annotate driving footage from months to days.

Labeling is the expensive part of building a self-driving system. Ordinary highway miles are easy to collect and teach the model little, whereas the rare events that matter must be found in the footage and manually annotated. A model that annotates its own data attacks that cost directly. Nvidia can commoditize the software layer because of the foundation it sits on. Automakers need NVIDIA clusters to fine-tune Alpamayo on their own fleet data, and DRIVE AGX is Nvidia’s in-car computer, optimized for it.

The customers are the carmakers and fleet operators who are not Tesla. Nvidia has named Lucid, Jaguar Land Rover, Uber, and the Berkeley DeepDrive research group as working with the Alpamayo family, and the Mercedes-Benz CLA is the first passenger car shipping with it on Nvidia’s DRIVE platform.

Bolt, Grab, Lyft, and Japan’s TIER IV are building robotaxi programs on the same hardware. Alpamayo now allows everyone a shortcut to developing self-driving, by skipping the step of funding a model from scratch.

All-In: Google’s AI Brain Drain, SpaceX’s Huge Quarter, Airtable’s 90% Collapse, US Data Fuels China AI