Tesla Files "Megapod" Trademark, Setting Up a Five-Layer AI Stack No Rival Owns

A June 18, 2026 USPTO filing for modular AI data center hardware reveals how Tesla plans to flip from compute buyer to compute seller—and eventually shrink the whole box down to a Powerwall bolted to your house that heats, powers, and pays you.

The most consequential Tesla product of the next decade may not be a car, a robot, or even the industrial AI crate the company just trademarked—it may be the extrapolation of that crate down to consumer size, a unit bolted to the side of a house that stores energy, runs private AI on Tesla silicon, heats the home with the waste heat that computing throws off, and rents its idle compute back to companies that pay by the month. That last idea isn't on any shelf and won't be for years. But every physical piece it requires already ships, and the business model behind it has already been described out loud. What makes it a when rather than an if is a single fact almost nobody has assembled: Tesla is quietly stacking five layers of the AI economy under one roof, and no other company on Earth—not NVIDIA, not Amazon, not Google—is even attempting all five.

Key Takeaways

• On June 18, 2026, Tesla filed USPTO trademark serial 99893717 for "Megapod," described as modular data center hardware systems for AI computing—an intent-to-use filing with no price or ship date yet.
• Megapod flips Tesla from compute buyer to compute seller; today its Cortex training cluster in Texas runs on roughly 67,000 NVIDIA GPUs the company rents, not builds.
• Tesla's AI5 chip taped out in April 2026, runs on about 250 watts, and is pitched as the best inference chip for edge compute—AI running in the field rather than in one central building.
• Five layers sit under one company: the chip (AI5), power (Megapack/Powerwall), cooling (15 years of thermal engineering), the fab (TeraFab), and the model (Grok and FSD).
• Roughly 25 gigawatt-hours of Megapack storage is headed to Italy and the UK via a June 23, 2026 Net Power deal—phase one of a program exceeding 100 GWh and north of $15 billion.
• Announced in March 2026, TeraFab—a joint chip fab with SpaceX, xAI, and later Intel—targets over a terawatt of AI compute a year, something NVIDIA and Apple don't attempt.
• Selling inference is already the cash engine: Anthropic pays xAI about $1.25 billion a month for Colossus compute, a contract worth over $40 billion, with Google signed near $900 million a month.
• Waste-heat computing already ships in Europe—UK firm Heata heats ~80% of a home's hot water off a server, and Microsoft pipes data center heat into a Danish town—proving the home box's core physics.

---

Dibs on a Name, Claim on a Business

An intent-to-use trademark is legal shorthand for calling dibs. Tesla isn't selling a Megapod, hasn't priced one, and hasn't dated one. What it has done is describe, in the plain language of a federal filing, a complete AI computer in a crate: servers, processing hardware, networking, power distribution, and cooling in one box you wheel in and plug into power. That specificity matters more than the absence of a product. Companies don't file for names on things they haven't scoped. The filing is the paper trail of intent, and it points in exactly one direction.

The tell is what the filing names as first-class components. Cooling being written into the product description isn't incidental—it signals that Tesla is treating thermal management as part of the box, not a bolt-on. That's the difference between a company selling chips and a company selling a system.

From Renting the Apartment to Owning the Building

Right now Tesla is a tenant. Its Cortex training cluster leans on roughly 67,000 NVIDIA GPUs—silicon Tesla pays for like everyone else. Megapod inverts that posture entirely. The analogy that fits is renting versus being the landlord: today Tesla writes the check, tomorrow it cashes them. A compute buyer is exposed to every price move its supplier makes. A compute seller sets the price.

That inversion is the whole strategic point. The value in AI doesn't accrue to whoever uses the most compute—it accrues to whoever controls the cost of producing it. Megapod is the vehicle for moving Tesla from the wrong side of that equation to the right one.

Why Inference at the Edge Is the Real Prize

Training a model is teaching it—the giant, one-time, warehouse-of-GPUs expense. Inference is using it, and it happens billions of times a day, forever, every time anyone asks an AI a question. Almost all the durable demand lives there. AI5 is built for exactly this: cheap inference running out in the world—in a car, on a wall, at a supercharger—rather than in one central building.

The power figure is the quiet bombshell. An AI5 chip sips roughly 250 watts, two or three light bulbs' worth. That number is what makes a consumer-scale AI unit physically plausible. You cannot bolt a warehouse of 700-watt datacenter GPUs to a house. You can bolt something that runs on the power of a few light bulbs, fed by a battery that's already there.

The Five Layers Almost No One Owns

Layer one is the chip: AI5, taped out in April 2026, roughly two years late, which is close enough to on time by Musk standards. Layer two is power—Tesla's energy division is among the largest battery deployers on the planet, and the June 23, 2026 Net Power deal for 25 GWh across Italy and the UK is only phase one of a program north of 100 GWh and $15 billion. Layer three is cooling, the actual bottleneck in AI today, backed by 15 years of thermal engineering on batteries and motors.

Layer four is the fab—TeraFab, announced March 2026 with SpaceX and xAI and later Intel, aiming past a terawatt of compute a year. This is the audacious one: NVIDIA doesn't own a fab, Apple doesn't own a fab, everyone rents from TSMC. Layer five is the model itself, Grok and FSD, so the brain isn't rented either. Line them up—chip, power, cooling, fab, model—and the vertical column belongs to one company.

The One Card NVIDIA Doesn't Hold

The obvious objection is that NVIDIA already sells the box. It does—the GB200 NVL72 is a pre-built, liquid-cooled rack of 72 top GPUs you wheel in and run. The turnkey AI crate isn't Tesla's invention; NVIDIA got there first and is extraordinary at it. But NVIDIA owns one layer—the chip and the rack around it—and then buys its power from the utility and rents its fab from TSMC.

That's the whole game. Owning one layer brilliantly makes you rich and dependent. Owning all five makes you the price-setter for the entire stack. Every layer NVIDIA rents is a cost Tesla can control internally, and cost is where all the money eventually comes from. No one—not NVIDIA, not Amazon, not Google—holds the full column.

The Cash Engine Is Already Running

This isn't a someday business. The Musk world already sells inference at scale that strains belief: Anthropic pays xAI roughly $1.25 billion a month—a month—for Colossus compute in Memphis, a contract worth over $40 billion across its life, with Google signed at something near $900 million a month. The landlord model is proven, except the tenants pay a billion dollars monthly.

Follow that cash and the strategy resolves. Inference revenue funds the fab, funds Optimus, funds the whole machine. Treating Tesla and SpaceX as separate companies misses the point—they're one combined compute apparatus, and selling inference to outside customers is becoming its beating heart. Megapod aims Tesla directly at that flow.

Shrinking the Crate to a Powerwall

The industrial Megapod is big and sold to companies. But the same five-layer stack already has a consumer form factor on millions of walls: the Powerwall. Today it stores solar and grid energy and sells power back. Put an AI5 chip inside it and it stops being a battery and becomes a tiny inference supercomputer running on 250 watts, fed by the battery already in the box, cooled by thermal knowhow Tesla already owns. Add as many chips or units as you like.

Musk has already sketched the distributed version of this—Digital Optimus units at superchargers, where Tesla has roughly seven gigawatts sitting idle, and the vision of parked Teslas pooled into a distributed AWS approaching 100 gigawatts of compute. The home box is simply the next logical node in a network the company is already describing.

The Box That Heats Your Home and Pays You Back

When a computer runs, its electricity becomes heat—that's physics, and data centers spend fortunes throwing it away. What if a house drank it instead? This already exists in Europe: the UK's Heata bolts a server to a hot water tank and covers about 80% of household hot water; French and German firms call them data furnaces; Microsoft pipes datacenter heat into a Danish town today. A computer that heats your home is a product you can already buy.

Fold it all into one Tesla unit and the box powers the house, computes for the house, and heats and cools it through an HVAC Tesla is well positioned to design given its history with cars and aircraft. The final piece is the payoff: when you're not using it, the box joins a distributed fleet and rents its spare compute and stored power to companies that need it, and the money lands in your account. Powerwalls already do a baby version through virtual power plants that pay owners for energy—except compute is worth far more than a few kilowatt-hours. Realistically, AI5 volume doesn't arrive until mid-to-late 2027, and early units might even ship with NVIDIA or older AI4 chips inside. But the direction is locked. The house that has cost money forever finally starts paying rent back.