On August 10, Fermi announced that TensorWave had signed a binding lease at Project Matador in Carson County — a 15-year deal Fermi values at about $6.5 billion, for a facility built to run large-scale AI training and inference on AMD Instinct GPUs, northeast of Amarillo in the Texas Panhandle. Shop owners who saw that headline usually do one of two things. They shrug: a GPU campus in the next county is not their problem. Or they wait: if industrial AI is coming here, maybe the phone line can stay as-is until the racks are hot.
Those are two different clocks. Mixing them up is how a shop sits out 2026.
Matador's clock
The shop's clock
Matador’s clock
Fermi America is building Project Matador as a power-first AI campus on thousands of acres in Carson County, northeast of Amarillo. The plan, Fermi says, ramps toward about 17 gigawatts, subject to customer agreements. Construction is underway. Fermi has said roughly 6 GW is already permitted, more than $1.5 billion is in the buildout, and first power is targeted in 2026. First power is not the same event as a tenant filling a hall with GPUs.
The TensorWave lease is the first binding customer agreement Fermi has announced for the campus. The parties are Fermi Campus 1 LLC and TensorWave TEX1, LLC, a subsidiary of TensorWave Inc. Fermi says the facility will have 222 megawatts of total facility power once the last delivery phase starts, and that it is designed to support tens of thousands of next-generation AMD Instinct GPUs for training and inference. Fermi puts contracted revenue at about $6.5 billion over the initial 15-year term for phase one, not counting two five-year renewal options. Phased delivery begins in the second half of 2027. The lease includes expansion rights for two more data centers that, if TensorWave exercises them, would take the partnership past 650 MW. Those are rights, not a second signed building.
TensorWave is an AI cloud that runs on AMD GPUs. It sells training and inference capacity to people who need a lot of it. It is not a receptionist for a plumber, and it is not a texting app for an Amarillo shop.
Training means teaching a model. Inference means using one — asking it to draft, classify, or answer. At Matador, inference is industrial: tens of thousands of GPUs, a private-grid power story, a 15-year clock. In a shop, inference is smaller and closer: a model suggesting the next text in a thread you already have with a customer. Same word. Different machine.
This is the first announced tenant after a messy year. Local and trade press covered a withdrawn earlier customer, water and leadership controversy, and a class action. The lease is binding and still has conditions Fermi and the press have described — project financing, guarantees, and Fermi’s note that some TensorWave obligations may be backed by an unnamed “global leader in AI.” We are not guessing who. Delivery of this facility starts in the second half of 2027, not this week. The lease is real. The calendar is still next year’s.
The shop’s clock
Your clock is the missed call this afternoon. The invoice that crossed two weeks. The voice note on the dash that was supposed to become an office update and never did.
None of that waits on Carson County. In the pattern post we named the shape: catch the trigger, write the honest checklist, pick a minimum viable robot, then reality-check it against a phone on LTE. We walked that through the text-thread payment nudge and AI-assisted drafting on a short leash. The failure modes are ordinary. Homeowners who hit voicemail call the next listing. Open invoices sit in the “I’ll get to it” pile until someone remembers to chase them. A tech finishes a job and the follow-ups live in someone’s head until Friday.
Before: the office manager reconstructs the week from memory on Friday, one customer never got the confirmation text, and the HVAC lead who hit voicemail at 2pm already booked the next name on the list. After: the missed call gets a text while the crew is still on the roof, the overdue invoice nudge is a two-tap approve, and the post-job note is captured before the next driveway. Same shop. Same people. A shorter gap between the work and the follow-up.
The robot that helps here is not a gigawatt campus. It is a text back when the call rings out, a payment link in the same thread the customer already uses, a draft the human still has to accept. Draft is not send. The leash is the product.
Where the clocks meet
When Matador is full of GPUs, there will be more industrial inference in the Panhandle in the abstract. That is a real story for power, land, and whoever rents those racks. It is not a reason to leave the shop phone on voicemail through 2026.
A 2027 campus will not catch a 2026 lead. Inference that matters to an owner-led shop is the kind that fits in a text thread, on one bar of signal, with a person still on the hook for tone and for whether the thing should go out at all.
Do not wait on Fermi to do the week’s work.
What to do this year
Three moves you can make without a 222 MW hall:
- Missed-call text-back. When you cannot pick up, the caller gets a text from the business number and a chance to stay in the thread instead of dialing the next name on Google. That is a safety net, not a substitute for answering — and it does not require a data center in Carson County.
- A payment nudge in the thread. When an invoice ages past the line you set, the office gets a prompt they can approve; the customer gets a link they can tap. The checklist is in the payment-nudge walkthrough.
- Short-leash drafts after the job. Capture the intent while it is fresh; let a model propose the office note, the invoice, the confirmation text; accept or discard per action. Nothing sends until a human says so. That walkthrough is AI-assisted drafting on a short leash.
If your shop is still running on voicemail, sticky notes, and the hope that someone checks the books this week: let’s talk. One conversation, zero obligation. We map the trigger and sketch what “fixed” looks like on this year’s clock — including missed-call work for trades and AI wired into the systems you already run.