Fing

Tesla fan 🛞 Food lover 🥗 Nutrition and Technology enthusiast 🍲

[Fing @Fing1907]

No matter how busy you are, remember this: Health is the top most important thing in life 🙏

[Fing @Fing1907]

@tslaming Yeah, this tracks 👀

Premiums are already booked out to ’27, so ditching the $41,990 AWD makes sense from Tesla’s end.

Still a pretty big jump tho for anyone who wanted that cheaper dual-motor, especially up in snow states or trying to stay under the NY rebate cutoff.

[Ming @tslaming]

SOS 🇺🇸 Tesla just pulled the $41,990 Model Y AWD from its U.S. order page, only about nine months after it launched 🛑

That’s a big deal if you need all-wheel drive. The cheapest AWD Model Y you can order now is the Premium at $49,990, so the price of getting into one just jumped by $8,000.

This trim was the sweet spot for snow-belt buyers. It took the base car’s battery, added a front motor, and ran 0-60 in 4.6 seconds with 294 miles of range. All of that cost only $2,000 more than the RWD.

New York buyers feel it too. At $41,990, it sat $10 under the $42,000 MSRP line where the state’s Drive Clean Rebate drops from up to $2,000 to just $500, and every AWD Model Y left on the page is priced above that line.

This doesn’t look like a demand problem. Every Model Y Premium is sold out for 2026, and new orders are now landing in January to March 2027. Back in late September, those same trims were still quoting October to November.

The Model Y L got pushed back too, now to February through April 2027.

Gas still costs about a dollar a gallon more than it did a year ago, buyers keep coming, and Tesla delivered about 22,000 more cars than it built in Q3.

With demand like that, every slot on the line probably earns more as a Premium than as a $41,990 entry trim.

Funny enough, this isn’t the first $41,990 Model Y to vanish. Back in early 2021, a Standard Range Model Y launched at that exact price and was gone within weeks.

This time, though, it’s heading out the door with Model Y orders already stretching into spring 2027.

[Fing @Fing1907]

@tslaming This is a really elegant solution 😍

Removing the actuators from the rack and solving tray clearance with a simple retractable plate keeps the high-speed links short while making swaps far less disruptive 🆒

[Ming @tslaming]

GOOD NEWS 🚨 Tesla just patented a clever way to wire supercomputer cabinets together where no hand can reach 🖐️

Inside Tesla’s AI training cabinets, six compute tiles sit side by side on a single tray. Each tile is an entire wafer of chips with built-in cooling, and each one talks directly to a partner tile in the cabinet next door.

Because high-speed signals degrade fast over distance, those links have to stay ultra-short. Tesla packs the cabinets shoulder to shoulder, running bundles of flexible cables straight through cutouts in the walls to plug into the tiles from underneath.

Once a tray locks into place, those plugs are completely out of reach—known in hardware engineering as blind connectors. Traditional setups rely on built-in actuators to line them up, which drives up part counts and burns valuable rack space.

And there’s a second headache: a connector sitting directly beneath a tile blocks the tray’s path. Pull a tray out for maintenance, and you risk smashing right into the very hardware feeding it data.

🔧 TESLA’S FIX: ONE TOOL DOES ALL THE HEAVY LIFTING

Instead of cramming motors inside the rack, Tesla stripped every moving actuator out of the cabinet and handed the job to a single manual tool that slides in alongside the trays.

Every cabinet-to-cabinet link consists of two mirror-image halves—one in each cabinet—bridged by ten flexible cables. On each side, the connector assembly rides on top of a long square tube running right beneath the tray, pointing straight up at the tiles. One tube holds three connector halves, meaning two tubes cover an entire six-tile tray, and engineers can easily stack more tubes higher up for multi-tray racks.

Each connector half is built from a stack of three plates, with each plate controlling a single axis of motion: the bottom plate lifts the plugs up into the tile, the middle plate nudges them side to side, and the top plate—which actually houses the connectors—slides them in and out of the tray’s path.

The tool itself is a long channel with a crank handle on one end that slides directly into the square tube. Spring-loaded bumps on either side snap into matching holes along the tube at each connector station, giving a distinct tactile feel and an audible click.

Give it a firm push to pop the bumps free, slide to the next click, and you’re at the next connector. A technician can pinpoint the exact link blindfolded just by counting clicks, like counting stairs in the dark.

Cranking the handle drives a push block up angled ramps inside the tool. As the block climbs, it presses against a knob under the bottom plate and hoists the connector straight into the tile’s socket—just like cranking a scissor jack to lift a car.

Two smart touches make that blind connection super forgiving: magnetic guides snap onto matching zones on the tray for rough alignment, while tiny springs let each connector float a hair in any direction so it settles smoothly into its socket.

Disconnecting works the exact same way in reverse, using a torsion-spring catch on the push block that hooks that same knob and pulls the connector down—a feature that can be integrated right into the primary tool so one tool handles both jobs.

With two ends to mate, your first instinct might be to plug both sides in at once. Tesla went the opposite route. Each half locks in independently from its own cabinet’s tube, making alignment way easier and cutting the required insertion force roughly in half.

Because the connecting cables are flexible, a tech can adjust or unplug one side while the other stays locked into the neighboring cabinet—even if the two racks aren't sitting perfectly flush.

That still leaves the tray-clearance problem, and Tesla solved it right inside the top plate. An angled cam slot is machined into the plate with a pin riding inside it.

When the tool pushes the pin from one end of the slot to the other, the slot pulls the entire top plate backward. The connector tucks neatly out of the tray’s way—like someone stepping back from the curb to let a bus pull in.

Now, every hardware swap follows a clean, repeatable sequence: the connector waits tucked back, the tray glides in, and the tool slides the connector forward under the tile before lifting it home.

Pulling a tray simply runs the choreography in reverse—drop, retract, and slide out—all while the other half of the cable link stays live and locked into the cabinet next door.

��� WHY IT MATTERS: DENSER RACKS, ZERO-DRAMA REPAIRS

That step-back mechanism lets Tesla pack cabinets tighter without sacrificing serviceability. Because the connector only enters the tray’s lane after the tray is seated, Tesla doesn't have to waste internal space on extra clearance—keeping cabinet-to-cabinet links short enough to preserve signal integrity.

Moving the mechanics out to a handheld tool also slashes internal complexity. With zero built-in actuators, there’s less hardware to manufacture, more room for compute, and Tesla can keep upgrading the tool long after the cabinets hit the data center floor.

Better yet, repairs stay completely localized. Techs can pull a tray from one cabinet without touching the rack beside it, and the remaining tiles can automatically reroute traffic around a dead neighbor so the cluster stays online.

When you're training massive neural networks for Autopilot and Full Self-Driving, keeping a cluster running right through a hardware swap is a huge deal.

While the wafer-scale tiles in the patent strongly resemble the original Dojo design Tesla shelved in 2025—and its latest roadmap shifts Dojo 3 toward chiplets with stacked high-bandwidth memory (HBM)—none of this mechanical wizardry actually depends on the silicon inside. The click-indexed tool and retractable plate live entirely at the cabinet level, meaning this exact setup can carry straight over to whatever racks Dojo 3 eventually fills.

Scaling an AI supercomputer means wiring endless rows of cabinets together, punching through another wall every time. Tesla just patented a clean way across—powered by nothing more than a tool in a tube and a connector polite enough to step aside.

[Fing @Fing1907]

This is a really clean solution 😍

By figuring out the turn a couple of nodes early and sending just a couple of lightweight flags ahead on the fast wires, the packet can keep moving without the usual stall at every hop. T

hose kinds of small wins in the interconnect start to matter a lot once the mesh gets big enough for serious training clusters 🔥

[Ming @tslaming]

BOOM 🚨 Tesla has patented a brilliant hardware trick that lets data blast across its AI chips without ever stopping to ask for directions 🚦

Tesla lays out its AI chips like a miniature city grid. A single silicon die can pack roughly a hundred compute blocks, called nodes, each about a square millimeter in size, with data packets darting between them.

At a 2 GHz clock speed, a packet gets just half a nanosecond to cross each node. That's a single clock cycle, the heartbeat keeping the entire chip in sync. Within that microscopic sliver of time, the hardware also has to lock in the packet's next move: straight ahead, a turn onto a cross street, or a stop at its final destination.

The conventional approach waits for a packet to arrive at a node before making a routing decision. One clock cycle is simply too tight to handle both tasks, so chips built that way end up burning extra cycles at every hop unless engineers shrink the nodes further than they want.

Those lost fractions of a nanosecond add up fast because node-to-node latency can bottleneck the entire chip. As long as every node has to receive a packet and compute its route on the same tick, the clock itself puts a hard ceiling on how fast data can move across the mesh.

🛣️ TESLA'S SOLUTION: DECIDING TWO BLOCKS EARLY

Tesla's breakthrough gets the routing decision out of the packet's way altogether.

It starts the moment a packet enters the grid. The chip checks a routing table to pinpoint which row or column the packet needs to turn on, or, for on-die trips, reads the target row and column straight from the packet itself. It even knows whether the packet will stop at a node after turning or exit at the edge of the die.

From there, each node looks two blocks down the road. If an eastbound packet has to turn at column 15, the node at column 13 already has the heads-up, just like a highway exit sign posted two miles ahead so drivers never have to slam on the brakes at the off-ramp.

Two tiny control signals carry that plan forward. A bypass flag, BYP, tells the hardware whether the packet can fly straight through the next node, while a "bypass next" flag, BYP_NEXT, signals the same thing for the node after that.

At every hop, the incoming BYP_NEXT shifts into the new BYP, and the current node calculates a fresh BYP_NEXT by comparing its own coordinates against the target turn. Those decisions roll forward like items on a conveyor belt, staying two steps ahead of the packet.

Here's how a single packet makes the run.

1️⃣ The first node receives both flags set to 1, letting the packet blow straight past that node and the next one.
2️⃣ As the packet exits, BYP stays at 1 while BYP_NEXT flips to 0 because the turn is coming up at the third node.
3️⃣ The third node sees BYP at 0 and steers the packet down its column.

That turn costs a little extra time. The node latches the packet into a flip-flop, a single-cycle storage element, and launches it onto the perpendicular track on the next tick.

Because the logic for east-west and north-south traffic often sits in separate regions of the die, a turn takes two clock cycles, so it pays to keep turns to a minimum.

The packet itself travels like a convoy with a few built-in tricks.

🚗 A lean header of roughly 20 bits leads the charge, while the 200-to-800-bit data payload trails one cycle and one node behind, steered by the same control signals captured in a flip-flop.
🚚 Heavy payloads can be sliced into two or four sub-packets, each carrying a duplicate header, with all of them moving in lockstep.
🔍 Those duplicate headers double as a parity check, so any mismatch at the next node trips an error flag, allowing the chip to reset, throttle down its clock, or bump up its voltage.
🚦 Bypassing isn't automatic. If packets are piling up or the queue ahead is nearly full, the node can prioritize other traffic instead.

So how does a flag computed at column 13 outrun a packet to column 15?

The secret lies in the chip's wiring, which stacks ten or more metal layers like floors in a skyscraper. The lower layers pack ultra-thin, dense traces for short local hops, while the upper layers carry thicker, wider wires that move signals faster.

Those top-floor wires are prime real estate. A single wide wire can eat up the footprint of three to five standard ones, so space is tight, and the natural instinct is to save them for heavy data.Tesla does the opposite.

The top one or two metal layers carry the bypass flags, while the layers right beneath them carry the packet payloads. The bulky data rides the standard, slower wires, and the tiny routing decision takes the express lane because the switch needs its instructions before the payload shows up.

Back at that turn, the flag shoots out of column 14 along a fast upper-layer wire and hits column 15 ahead of the packet. By the time the data arrives, the multiplexer, the hardware switch that selects the packet's outgoing port, is already locked in.

Looking two blocks ahead is just the baseline. Nodes can stretch that lookahead to three or four hops, buying time to steer packets around traffic jams on the fly, though every extra hop of foresight requires more of those scarce express wires.

That's where the old speed limit shatters.

Because the routing decision is already waiting when the packet arrives, a node no longer has to receive and decide in the same clock cycle. On a straightaway, the packet rides a dedicated through-path with no turns and clears each full-size node in a single tick.

🧠 WHAT IT MEANS FOR TESLA: FROM DOJO TO DOJO3

This idea comes straight from Tesla's silicon engineering team, and its first payoff is a trade-off Tesla no longer has to make. Without lookahead, designers are forced to choose between smaller nodes and slower hops. This design lets a chip skip that compromise entirely.

That's exactly the kind of chip D1 is. The custom processor behind Dojo, the supercomputer Tesla built to train FSD on real-world fleet video, packs its cores into an 18-by-20 mesh running at 2 GHz. The patent's inventor, Douglas Williams, is credited as one of D1's architects, meaning this routing scheme may have helped shape it from the ground up.

Speed is the next big payoff. In AI training, nodes work together on a single job and constantly pass results back and forth, so one slow hop can hold up the entire chip.

The difference jumps right off the page on a 10-by-10 grid. A corner-to-corner trip takes 18 hops with one turn, which works out to about 19 cycles, or under 10 nanoseconds. If every node had to stall for an extra cycle to compute the route, that same trip would balloon to around 37 cycles, nearly double.

Scale is where this approach pays off most. The design stretches to systems built across an entire silicon wafer, where grids grow massive and trips get longer, multiplying every cycle saved along the way.

The duplicate-header check matters at that scale too. It lets a giant system catch a garbled transfer and respond by slowing its clock or raising its voltage, rather than quietly passing bad data along.

Tesla can only put all of this to work on chips it designs in-house, which puts Dojo3 right at the center of the story. Tesla shelved Dojo in 2025, with Elon calling Dojo 2 an evolutionary dead end, before reviving Dojo3 in January 2026.

The company has said it's continuing to develop Dojo 3 to drive down AI training costs over time, and the stakes keep climbing. With its Q3 2026 Update putting training compute on track for 400 MW by year-end, every cycle a node spends waiting is compute Tesla pays for without using.

Cutting down on that wait time feeds straight into its cost goal. Cheaper training stretches every dollar further for the models powering FSD in owners' cars today, with Optimus up next.

For all the heavy data streaming through Tesla's AI chips, the real speed comes from something almost weightless. A single yes-or-no bit races ahead in the express lane, and by the time the data rolls in, the road is already clear. As these grids scale toward an entire wafer, that tiny bit may end up doing some of the heaviest lifting in Tesla's AI.

[Fing @Fing1907]

@tslaming Historic potential 🔥 Uniting EVs, robotics, AI, and space under one roof would create unmatched vertical integration 🔥

[Ming @tslaming]

MERGER NEWS 🚨 SpaceX and Tesla are moving closer to a merger, and a deal could be announced as early as December, according to the latest report from Bloomberg 🔥

Talks have shifted to deal terms. Boards and advisors on both sides are negotiating price and control, and one of the biggest open questions is how many SpaceX shares each Tesla shareholder would receive. That suggests SpaceX would absorb Tesla, rather than the other way around.

Bank of America analyst Alexander Perry expects an all-stock deal, valuing one Tesla share at roughly 2.8 SpaceX shares before any premium.

China is the biggest hurdle. SpaceX holds U.S. defense and government contracts, while Giga Shanghai produces more than half of Tesla’s vehicles, so the two sides are exploring ways to ring-fence Tesla’s China operations.

A December timeline isn’t set in stone, either. The companies have also discussed the announcement slipping into early 2027, and any deal would still require shareholder and regulatory approval.

There is no merger vote on the ballot for Tesla’s December 3 annual meeting, meaning any vote on a deal would happen later.

If the merger goes through, Musk’s EVs, robots, rockets, and AI would all sit under one roof, alongside Terafab, the chip venture the two companies already share.

[Fing @Fing1907]

This is the kind of scale that actually moves the needle 🔥

SpaceX taking the same manufacturing intensity that built Starship and applying it to AI infrastructure—while cleaning up a landfill, expanding water capacity, and aiming to lower local power bills—is a rare combination of ambition and local upside.

The space-computing angle makes it even more interesting. Huge if they pull it off.

[Ming @tslaming]

GOOD NEWS 🗞️ SpaceX is going all in on AI with more than $74 BILLION of new supercomputing campuses planned across Texas and Arizona.

The same machine that builds Starship and Starlink is now building supercomputers.

🚀 Starbase gets up to $24 BILLION for a facility near the Port of Brownsville that will train frontier AI and test tech for computing in space. It will make its own electricity instead of leaning on the public grid, and it has already created more than 3,000 construction jobs, with 500+ permanent jobs to follow.

🌵 Cochise County, Arizona, could land the biggest piece: a roughly $50 BILLION campus with about 5,500 construction jobs and 1,000 permanent ones. It would run on dry cooling with no planned industrial groundwater use, and SpaceX says it could cut local co-op members' power bills by up to 20%.

💧 Back in Texas, SpaceX is pouring $220 MILLION into Brownsville's water system to add about 20 MILLION gallons per day of new capacity, and $40 million has already landed. The city keeps full ownership, and SpaceX pays the same rates as everyone else.

🧹 Even the 444-acre Starbase site is getting a rescue, with decades of illegal dumping, thousands of tires included, being hauled away.

The scale here is wild.

SpaceX has already sunk more than $15 billion into Starship at Starbase, and it says the buildout ahead could dwarf all of it.

[Fing @Fing1907]

@tslaming Well said, Ming 👍 Nobody except Jensen should lead Nvidia and I can’t think of a better name right now 😅

[Ming @tslaming]

Personally, I don't think Tim Cook would be a good choice to lead Nvidia, even with President Trump's endorsement 💀

To be fair, Cook is one of the best operators of our time. He took over Apple in 2011, built a global supply chain nobody else can match, and grew it from a $350 billion company into a $4 trillion one before handing the CEO job to John Ternus on September 1.

But running Nvidia is a different job.

Nvidia already has elite operators. Colette Kress, now the interim CEO, helped steer it through its explosive growth.

What made Nvidia special was Jensen Huang's habit of betting on ideas years before anyone else believed in them.

CUDA launched in 2006, and when deep learning broke out in 2012, the model that kicked it off was trained on Nvidia gaming GPUs.

Now look at Cook's record on the two biggest tech bets of his era.

Back in 2017, Cook called autonomy "the mother of all AI projects." Apple's car effort ran for about a decade, cost billions of dollars and had roughly 2,000 people working on it by the end. In 2024, Apple shut it down without ever shipping a car.

On AI, Apple showed off a smarter Siri in June 2024, kept delaying it, and ended up turning to Google's Gemini to power it. Apple had to borrow Siri's brain from a rival.

His product list tells the same story.

Apple Watch and AirPods became huge businesses, but both grew out of the iPhone. Even Apple Silicon, the best thing to come out of his era, traces back to Apple buying chip startup P.A. Semi in 2008, while Jobs was still in charge. Vision Pro was his one truly new device, and it has stayed a niche product.

Cook spent 15 years harvesting what Steve Jobs planted, and he did it brilliantly.

But Nvidia doesn't need a harvester right now. It needs someone who can plant the next CUDA.

[Fing @Fing1907]

@tslaming Hmm, interesting, that basically confirms Nvidia will no longer be an innovative company 💀

[Ming @tslaming]

BOOM 🚨 President Trump just endorsed Tim Cook to be the next CEO of $NVDA 🔥

[Fing @Fing1907]

@tslaming @elonmusk let’s make it happen 🔥

[Ming @tslaming]

$TSLA and $SPCX are both green on the prospect of Jensen Huang’s leading the Terafab project 🔥 Now, just waiting for an announcement from Elon Musk 🔥

[Fing @Fing1907]

@tslaming That sucks 😱

[Ming @tslaming]

BOOM 🚨 Nvidia CEO Jensen Huang has resigned following revelations that he had two extramarital affairs, plunging the chipmaker at the heart of the global AI boom into crisis 🔥

[Fing @Fing1907]

This is a really elegant solution 😍

By turning the cathode powder into its own self-supporting saggar with built-in channels, Tesla solves the density, heat and gas transfer, and consumable cost problems all at once.

It feels like a smart way to get more throughput out of the most expensive step in the process.

[Ming @tslaming]

BATTERY NEWS 🔋 Tesla has patented battery bricks that bake without a box, then get smashed on purpose 🔥

The costliest step in making cathode material, the metal-rich powder on a battery’s positive side, looks a lot like baking. The powder rides through a kiln for more than 12 hours inside ceramic boxes that typically need replacing after just 1 or 2 weeks.

Cathode makers call those boxes saggars, and they carry a lithium compound mixed with a metal precursor through a long tunnel oven called a roller hearth kiln. The heat reacts the mix into finished cathode in a step known as calcination. No stage of cathode making costs more, because the kilns are the priciest machines in the line, burn the most energy and hold material the longest.

The box itself slows everything down. Powder sits still inside it, so heat and gas move through slowly and every batch needs more time in the kiln. The exit end has to cool gently so the saggars don’t crack, and the machines that handle and inspect them are costly and cause frequent downtime.

Each box also has a hard ceiling. Loose powder weighs just 0.9 grams per cubic centimeter (g/cm³), lighter than water, and a saggar normally holds only about 4.5 kg of it. Loading more can hurt quality by disrupting gas flow and heat spread, and stacking boxes, commonly 4 across and 2 high, doesn’t scale.

🧱 TESLA’S SOLUTION: A BRICK BUILT TO BE BROKEN

Tesla’s answer is to throw away the box and make the powder hold its own shape.

The process starts in a mixer, where a lithium reagent such as lithium hydroxide or lithium carbonate meets a metal precursor. That precursor is a hydroxide, oxide or carbonate of metals like nickel, manganese, cobalt or aluminum. Then Tesla stirs in up to 20% water and a pinch of binder, a polymer glue such as PVA or CMC, at 1% or less.

Next, a press squeezes the damp mix into a brick or a tile. Tesla packs it to 1.9 to 2.3 g/cm³, more than double the density of the loose powder. The result is a self-standing element, one that keeps its shape under its own weight with nothing around it.

Squeezing alone isn’t enough. Tesla pressed one brick dry, with no water or binder, to 1.7 g/cm³, and over time it crumbled. A brick made with water and binder at about 1.8 g/cm³ held its form, the way damp sand holds a castle that dry sand can’t.

Packing the powder twice as tight sounds like it should make baking worse. Deeper powder already hurts gas flow in a saggar, so a dense block looks like it would trap its own core, cut off from the oxygen the reaction needs.

Tesla’s fix is to build the airways into the brick itself. One design runs through-holes along the full length of the block, like the cores of a cinder block. Tesla’s tests used 12 of them, each 20 mm across, in both round and square versions.

The surface carries airways too. Rows of half-round grooves line the sides of the bricks, so neighbors form channels where they meet. Thinner tiles, 10 to 50 mm thick, get a wavy face and stack like sheets of corrugated cardboard, with tunnels between the ridges.

Each channel is an enclosed tunnel open at both ends, 5 to 100 mm wide and 10 to 200 mm from the next. During the bake, oxygen flows in to reach the reactants at the center, and water vapor from the reaction escapes. The rest of the brick stays tightly packed, and the open paths may also help keep it from cracking.

Now the bricks can travel bare. Tesla stacks them straight onto flat ceramic plates, several layers high, and can send them through a gentler pre-bake first. In one test, six stacked tiles rode through flowing hot air at 250 °C (482 °F).

That warm-up can drive off free water, pull the bound water out of lithium hydroxide and turn metal hydroxides into oxides. The binder burns or chars away as the heat builds, and the brick gets less dense as it goes. Any residue left behind can vaporize in the hotter stage that follows.

The main bake runs at about 650 to 850 °C (roughly 1,200 to 1,560 °F). Tesla picks an oxidizing, inert or reducing atmosphere to suit the chemistry, and it can pass gas straight through the holes and channels as the bricks cook. Nickel-rich recipes in Tesla’s tests baked for 12 hours in flowing oxygen.

The bricks handled a wide spread of chemistries in Tesla’s tests. A lithium manganese oxide (LMO) brick baked at 850 °C for 18 hours in air. Nickel-manganese-cobalt (NMC) bricks ranged from 50% to 80% nickel, and a nickel-cobalt-aluminum (NCA) brick rounded out the set.

The finished brick lands at 1.7 to 2.3 g/cm³, with at least 95% of its weight in cathode active material. Tesla built the process for improved crystallinity, atoms lined up in a tidier lattice, which tends to mean better battery performance. It can get there without an extra heating step, the kind of add-on that drags down the old way.

Not every recipe survived the trip. One tile mix used 0.25% CMC and 7% water, and a stack of six stayed standing through pre-bake and calcination. A second mix with the same ingredients but only 0.05% CMC and 3% water couldn’t stay standing, and that stack slumped into a heap laced with cracks.

A fraction of a percent of glue and a few points of water can separate a clean stack from a pile, and a pile keeps none of its tunnels. Tesla’s brick runs showed the same split, with some batches leaving the kiln whole and others torn by cracks.

Then Tesla destroys the brick it worked so hard to protect. Once the bricks cool, it flips the plates, crushes them, mills the chunks and sifts the powder through a fine 400-mesh screen. It then inspects the empty plates and loops them back to carry the next stack.

After that, Tesla can wash, filter and dry the powder, then fuse on a coating of just 0.3% nano aluminum oxide or 0.5% nano titanium dioxide by weight. The brick itself never reaches a battery. It was the box all along, made out of its own cargo.

🏭 WHAT IT MEANS FOR TESLA: MORE CATHODE ON EVERY TRIP

Every extra kilogram per load counts, because Tesla says batteries are the bottleneck. Its Q3 2026 update named battery supply as the limiter on how fast it ramps vehicles, with Texas cathode expansion among the fixes.

One brick shows how much room the box was wasting. Tesla’s common test brick measures 100 by 150 by 300 mm, and pressed solid to 1.8 g/cm³, it carries about 8 kg, nearly double a saggar’s ceiling.

It does that on less than half a saggar’s 330 by 330 mm floor, with room for more bricks stacked on top. Swapping each box for one solid brick means about 1.8 times the material per load, before counting the floor space saved.

That extra cathode would have somewhere to go. Since August, Austin has run what Tesla calls the first large-scale cathode plant in North America, built to feed its 4680 cells. A month later, the first Cybercab rolled out carrying that homegrown cathode.

Inside, it reportedly fires nickel, manganese and cobalt precursors with lithium hydroxide in a kiln, the same pairing several brick tests used. Tesla’s Gulf Coast refinery is built to make about 20,000 metric tons a year of lithium hydroxide monohydrate, the exact reagent in the winning tile recipe. These bricks could slot straight into a supply chain Tesla already runs.

For owners, the chemistry stays familiar. The NCA and NMC families Tesla baked as bricks already power many Teslas on the road, and the design covers LFP too. The first cells to feel the change could be 4680s, which Tesla is ramping for Cybercab, Semi and Model Y.

The old picture was fluffy powder creeping through a kiln in boxes that need replacing every 1 or 2 weeks. In Tesla’s design, the powder carries itself through, airways and all. The only thing that breaks at the end is the brick, and Tesla breaks it on purpose.