Tesla Files 'Megapod' Trademark, Setting the Stage for an AI Box That Pays Homeowners

A June 18th filing for modular AI data center hardware reveals how Tesla's five-layer stack — chip, power, cooling, fab, and model — points toward a home unit that computes, heats, powers, and earns.

Tesla's most obvious future product is one it hasn't announced, priced, or dated — yet every component already exists inside the company. A single trademark filing exposes the trajectory: Tesla is quietly turning itself from a buyer of compute into a seller of it, and the logical endpoint is a box bolted to the side of a house that runs private AI, heats the home with its own exhaust, powers the house from its battery, and rents out spare capacity for cash. The industrial version is real and filed. The consumer version is a short extrapolation from parts already shipping.

Key Takeaways

• On June 18th, 2026, Tesla filed a trademark (serial 99893717) for "Megapod," described as modular data center hardware systems for AI computing.
• Currently Tesla is a compute buyer, running its Cortex training cluster in Texas on roughly 67,000 NVIDIA GPUs it does not own.
• Tesla's AI5 chip taped out in April 2026, pitched as an edge inference part sipping about 250 watts — the draw of two or three light bulbs.
• Tesla Energy signed a June 23rd, 2026 deal with Net Power for 25 GWh of Megapack storage across Italy and the UK, part of a 100+ GWh, ~$15 billion program.
• TeraFab, a Texas chip fab jointly announced by Tesla, SpaceX, and xAI with Intel joining, targets over a terawatt of AI compute a year.
• Anthropic reportedly pays xAI $1.25 billion a month for compute on the Colossus supercomputer — a contract worth north of $40 billion over its life.
• Europe already sells waste-heat computing: UK firm Heata claims a server on a hot water tank delivers about 80% of a household's hot water.
• Because Tesla owns chip, power, cooling, fab, and model, it controls the full cost stack that NVIDIA, Amazon, and Google each rent from someone else.

---

The Filing That Flips Tesla's Role

An intent-to-use trademark is legal dibs, not a product — there is no Megapod to buy, no price, no ship date. But the description is precise: modular data center hardware for AI, meaning servers, processing silicon, networking, power distribution, and cooling packaged as one crate you wheel in, plug into power, and run. The significance isn't the box; it's the direction of the arrow.

Today Tesla writes checks to NVIDIA like everyone else. Megapod inverts that. It moves Tesla from renting the apartment to owning the building — from compute buyer to compute landlord. For a company whose energy and AI ambitions Wall Street chronically underrates, that inversion is the whole story.

Five Layers Under One Roof

The reason Tesla can attempt this when a dozen rivals would love to sell the same crate comes down to a stack almost no one else holds. Layer one is the chip: AI5, built for cheap inference at the edge. Layer two is power: Tesla Energy is among the largest battery deployers on Earth. Layer three is cooling — the actual bottleneck for modern AI — where Tesla brings 15 years of thermal engineering on cells and motors, and which the Megapod filing names as a first-class component.

Layer four is the factory: TeraFab, the attempt to own the building where silicon is etched, something NVIDIA and Apple both outsource to TSMC. Layer five is the model itself, via Grok and Tesla's self-driving stack. Line them up and the picture clicks: chip, power, cooling, fab, and model in one company. Cost is where all the money lives, and Tesla is the only firm reaching for control of every layer of it.

Inference, Not Training, Is Where the Demand Lives

The distinction that makes AI5 matter is between teaching a model and using one. Training is the giant, expensive, warehouse-scale event — the 67,000 GPUs. Inference is what happens every time a chatbot answers a question, billions of times a day, forever. That is where the durable demand sits, and it is what edge silicon is built to serve: computers placed out in the world rather than in one central hall.

At roughly 250 watts, AI5 is designed to run AI cheaply by the millions — in cars, at Superchargers, on walls. That's not a data-center part fighting for a river-cooled rack; it's a distributed inference engine. Once you see AI as a utility delivered at the edge, the home box stops looking exotic and starts looking inevitable.

Where NVIDIA's Lead Actually Ends

The obvious rebuttal is that NVIDIA already sells the turnkey box — the GB200 NVL72, 72 top GPUs wired into a liquid-cooled rack you roll into a warehouse. That's fair; the pre-built AI crate is not Tesla's invention, and NVIDIA is extraordinary at the one layer it owns. Jensen Huang's execution speaks for itself.

But NVIDIA sells the rack and then buys its power from the utility and rents its fab from TSMC. It owns the chip and the rack around it — one layer. Tesla is going for all five. No one, not NVIDIA, not Amazon, not Google, owns the entire vertical column. Owning it means controlling the cost of it, and that is the moat.

The Cash Engine Already Running

Selling inference is not a someday business for the Musk world — it's live at scale that strains belief. SpaceX now leases its Memphis supercomputer, Colossus, to outside customers. Anthropic reportedly pays xAI $1.25 billion a month, a contract worth over $40 billion across its life; Google signed something in the neighborhood of $900 million a month. The landlord model is proven, and the tenants pay in billions.

That cash flow is the engine funding the fab and the robots. The deeper insight is structural: Tesla and SpaceX aren't two separate companies but one combined compute machine, and selling inference to outsiders is becoming the cash-flow heart of the whole empire. Megapod points Tesla straight into that revenue stream.

From Powerwall to a Supercomputer on the Wall

The industrial Megapod is big and sold to companies, but the same five-layer stack already has a consumer shell installed on millions of homes: the Powerwall. Today it merely stores solar or grid energy and sells some back. Drop an AI5 chip inside and the battery becomes a tiny inference supercomputer on the wall — running on 250 watts fed by the pack that's already there, cooled by thermal know-how Tesla already owns. And the company can add as many chips, or as many Powerwalls, as it likes.

This isn't a leap; it's a substitution of one component into an enclosure that already ships. Tesla has even sketched the surrounding vision — Digital Optimus deploying inference units at Superchargers, where roughly seven gigawatts of power already sits, and pooling idle Teslas into a distributed answer to AWS, up to 100 gigawatts of compute doing nothing in driveways.

Waste Heat Becomes a Feature

When a computer runs, the electricity it burns becomes heat — pure physics, and normally pure waste that data centers spend fortunes discarding. Flip the frame and the waste becomes a household utility. This already exists in Europe: Heata in the UK bolts a server to a hot water tank and claims about 80% of a home's hot water from the exhaust; France and Germany run "data furnaces"; Microsoft pipes data-center heat into a Danish town's district heating.

Fold that into a single Tesla unit and the box does four jobs at once: it stores the home's electricity, runs its private AI, heats the house through an HVAC Tesla is well positioned to design itself, and cools it — all off a stack the company builds end to end. The heating isn't a gimmick bolted on; it's the natural byproduct of the compute you already wanted.

The Box That Pays You Back

The final piece is the one that changes the homeowner's relationship to the machine. When idle, the unit joins a distributed fleet and sells spare inference — or the energy in its batteries — to buyers who need it, and the money lands in your account. The Powerwall already runs a baby version through virtual power plants that pay people for energy sent back to the grid. The upgrade is that compute is worth far more than a few kilowatt-hours.

The timeline is honest: AI5 doesn't reach real volume until mid-to-late 2027, the fab is a 2027-and-beyond story that could slip, and the first home unit might even ship with NVIDIA or older AI4 silicon inside. But the direction is locked. Megapod is real and filed; the home box is a basic extrapolation built from parts that already exist and a business model already described aloud. For most people, the AI age may not arrive as a chatbot but as the appliance on the side of the house that finally starts paying them back.