Tesla Files Megapod Trademark for Modular AI Compute Hardware

A June 2026 USPTO filing for modular data-center systems, paired with AI5, energy storage, and a five-layer stack, points Tesla toward selling inference compute instead of only buying it.

Tesla just claimed the name for a product that does not ship yet and still reads like the logical end of everything the company already builds. Megapod is modular AI hardware in a crate - processing, networking, power, and cooling in one box - and the filing is an intent-to-use claim, not a price list or a ship date. Once you line up the chip, the batteries, the thermal work, the fab ambitions, and the models under one roof, the industrial box and a home-sized cousin stop looking speculative. They look like the same stack, scaled down to a wall.

Key Takeaways

• USPTO serial 99893717, filed June 18, 2026, covers modular data-center hardware for AI computing under the Megapod name.
• Cortex still trains on roughly 67,000 NVIDIA GPUs, so Tesla is a compute buyer today; Megapod is the flip to selling compute.
• AI5 taped out in April 2026, runs near 250 watts, and is aimed at edge inference rather than warehouse training.
• NatPower deal on June 23, 2026: 25 GWh of Megapack storage in Italy and the UK in phase one, with a full program above 100 GWh and north of $15 billion over its life.
• TerraFab, announced March 2026 with SpaceX and xAI (Intel later in), targets over a terawatt of AI compute a year - a fab path almost no chip buyer even attempts.
• Anthropic is paying xAI about $1.25 billion a month for compute on Colossus; a Google-scale deal sits near $900 million a month.
• NVIDIA's GB200 NVL72 already sells a liquid-cooled rack; Tesla is going for chip, power, cooling, fab, and model together.
• Digital Optimus and driveway cars sketch distributed edge inference at Superchargers (~7 GW of power on hand) and up to ~100 GW of parked compute.
• Home Powerwall-plus-chip logic is extrapolation, not a SKU: storage, private AI, waste heat for the house, and spare inference sold back into a fleet.

---

Megapod Is a Crate, Not a Slogan

The filing language is dry on purpose. Modular data-center hardware systems for artificial intelligence computing means servers, AI silicon, networking, power distribution, and cooling packaged so you roll the unit in, plug it into power, and run inference. There is no public price, no ship window, and no product page. Intent-to-use is the legal move for calling dibs on a name before the crate exists.

That still matters. Tesla's big training cluster, Cortex, sits on NVIDIA silicon. Writing that check is the industry default. Megapod is the other side of the ledger: Tesla as the party that owns the box and rents out what it does. Same company, inverted role. For anyone watching unit economics, buyer versus landlord is the whole game.

Five Layers Under One Roof

Why attempt a crate full of AI when plenty of vendors would love that revenue? Cost is the wall. Stack ownership is the answer. Tesla is assembling five layers almost nobody else holds together.

Layer one is the chip. AI5 finalized design and went to the fab in April 2026 - late if you live by calendar optimism, roughly on brand for long silicon schedules. The pitch is inference at the edge: not teaching a model in a warehouse of 67,000 GPUs, but answering queries out in the world, millions of times, for cheap. About 250 watts per chip is two or three light bulbs, which is the power budget you need if the computer lives next to a house or a Supercharger instead of next to a river.

Layer two is power. Energy is the division Wall Street still treats as a side quest. Megapacks and Powerwalls already put Tesla among the largest battery deployers on earth. The NatPower agreement alone - 25 GWh in phase one across Italy and the UK, full program past 100 GWh and more than $15 billion over its life - is more storage in an opening tranche than Tesla deployed globally across all of 2026 to date. That is not a hobby. That is a power company that also builds cars.

Layer three is cooling. Heat is the quiet bottleneck for AI racks. Data centers hug water and baseload for a reason. Fifteen years of battery and motor thermal work is not marketing copy when the filing names cooling as first-class product, not an afterthought bolted on later.

Layer four is the factory. TerraFab, the joint chip fab story with SpaceX and xAI in Texas, with Intel joining after the March 2026 announce, aims at more than a terawatt of AI compute per year. NVIDIA and Apple rent capacity at TSMC and a few U.S. sites. Owning the building where the silicon is etched is a 2027-and-beyond bet and can slip. Almost no peer is even trying.

Layer five is the model. Grok through xAI and Tesla's FSD stack mean the brain is not rented from someone else's API either. Chip, power, cooling, fab, model. Cost control lives in that column. Margin lives there too.

NVIDIA Already Sells a Box - And Still Rents the Rest

The fair pushback is simple. NVIDIA already ships turnkey AI iron. The GB200 NVL72 is a pre-built liquid-cooled rack: 72 top GPUs wired together, wheeled into a warehouse, ready to run. The crate idea is not original to Tesla. NVIDIA got there first and is excellent at that layer.

What NVIDIA does not own is the full vertical. It sells the rack, then buys power from utilities and rents fab space. One brilliant layer, executed at a scale that remade the market. Tesla is trying for all five. Amazon and Google run huge fleets and still do not hold chip design, home and grid batteries, car-grade thermal systems, a owned fab path, and frontier models in the same corporate stack the way this empire is lining up. Control of cost across the column is the thesis. Everything else is theater.

Compute Rent Is Already Paying the Machine

Selling inference is not a slide deck fantasy in this orbit. Colossus in Memphis already leases capacity. Anthropic's Claude stack is on the order of $1.25 billion a month for compute. Google-scale contracts sit near $900 million a month. That is landlord math with billion-dollar tenants. Cash flow of that size funds fabs, robots, and the next wave of iron. Treating Tesla and SpaceX as disconnected tickers misses the combined compute machine: train and serve at industrial scale, recycle revenue into the next layer, push inference toward the edge where the demand lives.

Edge First, Then the Wall

Digital Optimus sketches dedicated inference units at Superchargers, where roughly 7 GW of power already sits. Cars in driveways become a distributed pool - the AWS analogy, only scattered - with talk of up to 100 GW of idle silicon if the fleet is orchestrated. The industrial Megapod is the big sibling. The home box is the same five-layer logic, shrunk to a product Tesla already bolts to walls by the millions.

Powerwall today stores solar or grid energy and can sell power back. Drop AI silicon into that envelope and the box stops being only a battery. It becomes a small inference node on 250-watt sips, fed by its own pack, cooled with know-how already in the company. How many chips or how many walls is a product decision. The physics are not mysterious.

Waste Heat Stops Being Waste

Every watt a computer burns becomes heat. Data centers pay to throw that heat away. Homes already buy heat. In the UK, Hiata-style setups bolt servers to hot-water systems and cover on the order of 80% of household hot water from that exhaust. France and Germany run the same "data furnace" idea. Microsoft has piped data-center waste heat into district heating in Denmark. A computer that warms a building is a product category that already exists in Europe.

Fold storage, private home AI, waste-heat HVAC (cars and aerospace thermal work are not theoretical for this group), and grid or compute export into one unit and the house stops being only a monthly cost. The Powerwall virtual power plant is the baby version: Tesla already pays owners for energy sent back. Selling spare inference is the same model with a higher-value commodity than a few kilowatt-hours. That line is the payoff for the homeowner, not the warehouse operator.

Timeline Stays Honest

AI5 volume is a mid-to-late 2027 story. No home Megapod sits on a shelf. Early units could even ship with NVIDIA parts or today's automotive AI4 silicon before Tesla's own inference silicon is everywhere. Direction is clearer than date. The industrial trademark is real and filed. The wall unit is extrapolation built from parts that already ship and a distributed-inference story already described in public for cars and Superchargers. The AI age for a normal household may show up as the box on the side of the house - the one that used to only cost money - starting to pay something back.