Waymo's Custom Silicon Is a Real Step Down in Cost. At $125,000 a Car Against a Cybercab We Put at $20,000-$30,000, It Is No Threat Yet to Tesla's Scale.

X's news module carried it on Thursday as "Waymo Unveils Custom 5nm Chip Slashing Robotaxi Costs to $20,000" — 145 posts and climbing.

Start with the credit, because it is real: Waymo has taken $75,000 out of this robotaxi in one generation, and Thursday's chip is a serious piece of engineering. Then read the sentence again, because two things in it are wrong and both are checkable against the primary source.

Waymo's post contains no cost figure. "A look under our trunk: what's in our compute", published the same day, is an engineering post. No price, no bill of materials, no generation number. Neither did the 6th-generation Driver announcement in August 2024, which said only "significantly reduced the cost" and "a fraction of the cost". Waymo has never published a per-unit cost in either post. Every dollar figure in circulation is reported rather than disclosed.

And the flagship number is the front end, not the driver. Waymo's own wording:

To handle the massive influx of raw data before it reaches our core ML brain, we are excited to introduce our purpose-built 5nm ASIC… engineered exclusively to process, fuse, and run advanced neural networks on raw sensor data in real time… While these ASICs alone deliver over 1,000 TOPS of ML performance dedicated to front-end processing…

More than a thousand trillion operations a second, consumed before the driving starts — extracting and fusing raw lidar, radar and camera streams, including "temporal denoising for superior low-light perception". The output is handed to a separate inference engine, which is where the car decides anything. Waymo publishes no figure for the whole system. The number everyone is quoting measures how much silicon it takes to read Waymo's sensors. The cost figures attached to it in the coverage are for the driver hardware package — sensors and compute, bolted on top of a vehicle — and on Morgan Stanley's estimate of the finished car that package is only 16% to 20% of it.

What's in the new hardware

The chart above is the whole disclosure, as a list. Everything Waymo published:

• A custom 5nm ASIC, purpose-built, in production.
• Over 1,000 TOPS of ML performance from those ASICs — for front-end processing, not for driving.
• What that front end does: process, fuse and run neural networks on raw lidar, radar and camera data, including temporal denoising for low light, before handing off to a separate inference engine.
• The suite it reads: 13 cameras, four lidar, six radar, with 17-megapixel cameras.
• Two independent engines — compute duplicated, running full parallel workloads, either able to take over.
• Integration with the vehicle's liquid cooling system.
• 20x compute growth in eight years, about 45% a year compounded.
• Over 200 million miles of fully autonomous experience behind the design.
• Named partners: AMD, Micron, NVIDIA, Samsung, Sandisk, Socionext and TSMC.

And what it did not publish: any cost figure at all.

The sensor suite, rendered in silicon

This is where the announcement points in the opposite direction from the headline.

The 6th-generation Driver carries 13 cameras, four lidar and six radar — 23 sensors, disclosed by Waymo — plus an array of external audio receivers, with new 17-megapixel cameras for range and low light. That is already the streamlined version: the 5th-generation stack on the Jaguar I-Pace ran 29 cameras and five lidar, and the reduction is reported at 42% fewer sensors (43% by our count of the disclosed figures). Waymo's framing is that it "reduced the number of sensors while maintaining our safety-critical redundancies", and the payoff is real — better performance in snow and rain, which is what opens colder cities.

Thursday's post is what it costs to read what remains. Thirteen 17-megapixel camera streams fused in real time against four lidar and six radar, denoised for low light, at over 1,000 TOPS, on a custom 5-nanometre part. A front end large enough to justify its own silicon programme.

Tesla's stack does not have this stage. No lidar to denoise, no radar return to fuse against a camera frame, no cross-modal front end to build, because the sensors are not there. You can disagree with that decision on safety grounds; nobody disputes it on cost grounds.

Waymo then buys the redundancy twice:

Our compute is designed like two independent engines. While they normally operate as one unit running full parallel workloads, if one experiences a fault, the other seamlessly takes over.

That is the correct engineering answer when there is no human backup, and Waymo is right to build it. It is also hardware bought twice, plumbed into the vehicle's own liquid cooling loop. The picture is a driver package that is architecturally heavier at every layer by design: more sensors, more front-end compute to read them, duplicated, and cooled by the car.

None of this is a criticism of the engineering. Waymo says it has scaled onboard compute 20× in eight years — about 45% a year compounded — on a part that has to survive vibration, shock and a Phoenix summer. It is a description of what the cost target has to pay for.

The other tax the sensors levy: about a fifth of the battery, daily

Cost is the visible half of the sensor decision. Power is the half that lands on utilisation.

Waymo publishes no wattage. Thursday's post talks about "maximizing battery efficiency" and the liquid-cooling loop, and gives no number. The figure analysts consistently use for a full multi-modal Level 4 stack — sensors plus compute — is about 1 kW continuous, and it is theirs, not Waymo's.

Take it at face value against the Ojai's reported 93 kWh pack. A 20-hour operating day is 20 kWh; a 24-hour one is 24 kWh. That is 22% to 26% of the battery consumed before the vehicle moves a single passenger.

The electricity is nearly free — 7,300 kWh a year is a few hundred dollars, against $96,360 for one safety driver. What it costs is time on a charger instead of on a fare, and that lands on the one slider our own Tesla model says matters most in this business: "Utilisation matters more than fleet size. Doubling the fleet at 45% utilisation is worth less than holding the fleet and reaching 72%." A fifth of the pack spent on perception is a permanent haircut to the number the model is most sensitive to.

Tesla's draw is estimated lower — no sensors to power, less to process — but Tesla publishes no figure either, so there is no comparison here worth printing, only a direction.

Where the two vehicles can be set side by side is at the wheels, on reported specifications:

Tesla gets more range out of a battery 48% smaller. On a 200-mile day the Ojai draws 70–100 kWh against 33 kWh — two to three times the energy for the same work, on a pack nearly twice the size, which is itself real money in a vehicle whose cost is the problem.

Be careful about what that proves. Most of the gap is body, not sensors. A six-seat minivan is not a two-seat coupé, and no amount of silicon closes that. At roughly 1 kW and urban average speeds, the driver stack accounts for something like 33–50 Wh/mi — 10% to 27% of the gap. Put a Waymo-class sensor load on a Cybercab and it would still be near 205 Wh/mi. The sensors are the smaller half of this comparison too; the vehicle is the bigger one. Which is, again, the thing Thursday's announcement did not touch.

What is custom, and what is still bought

Worth noting who else is in the trunk. Waymo names its partners:

We are proud to work alongside a number of partners like AMD, Micron, NVIDIA, Samsung, Sandisk, Socionext, and TSMC…

So Nvidia and AMD supply into the same system whose new chip is reported as matching Nvidia's autonomous-driving hardware — a characterisation that appears in the trade coverage and not once in Waymo's post, which makes no competitive claim about anyone. Micron, Samsung and Sandisk are the memory names. Socionext is an ASIC design house, which is a fair hint at how a robotaxi company got a 5nm part designed at all. TSMC is named as a partner; that it fabricates this specific chip is press-reported rather than stated, and TSMC's own margin structure is something we looked at last week.

The custom part is the front end. The merchant silicon is still in the car.

The cost stack, with every layer labelled

Here is the comparison the headline invited, done with the provenance attached. The vehicle is the Ojai, purpose-built by Zeekr — the Geely-owned brand — and now open to all riders in San Francisco, Los Angeles and Phoenix, with roughly 300 units in commercial service.

Two independent build-ups bracket the same car: $78,000 of landed chassis plus $25,000 of driver hardware is $103,000, and Morgan Stanley's $125,000 leaves about $22,000 for fit-out and integration. They agree on the shape, and the range most commonly quoted for a finished Ojai — $110,000 to $125,000 — sits inside them. We use the top of it below, because it is the figure that flatters Waymo least badly and still makes the point.

Why we now put the Cybercab at $20,000–$30,000

Our published fleet math used $30,000, and that figure came from a Model-Y-shaped vehicle. The Cybercab is not one, and the disclosed attributes all point the same way:

• A 48 kWh pack against the Ojai's 93 kWh — 52% of the battery, and the pack is the single largest line in any EV bill of materials.
• Two seats, no steering wheel, no pedals, no mirrors, no instrument cluster — an interior with most of a normal car's parts deleted rather than redesigned.
• A purpose-built line at Giga Texas running Tesla's unboxed assembly process, at >125,000 units a year of installed capacity, which is the volume at which fixed cost per unit stops mattering.

That is a case for the low end, not a disclosure of it. Tesla has never published a Cybercab build cost, and we are not going to invent a point estimate for one. What we will do is widen our own number to $20,000–$30,000 and carry both ends through every comparison below, because the conclusion does not depend on which end you believe — and it is worth being explicit that the low end is the one that makes Waymo's position worse.

Now the part that matters, and it is not the chip.

Waymo took $75,000 out of this vehicle and it is the best thing in the story. A 37.5% cut per robotaxi, driver hardware down more than 50% a generation, sensor count down 42%. That is a company executing hard on cost, and anyone dismissing Thursday as marketing is not reading the trend line.

And the car is still four to six times a Cybercab. Against the $20,000–$30,000 range above — our number, not Tesla's — the Ojai is 4.2× to 6.3× on Morgan Stanley's figure, and 3.4× to 5.2× on the import build-up. Across the 14,800-car fleet our robotaxi fleet math identified as 1% of Tesla's revenue, that is $1,850M of hardware against $296M–$444M.

The chip cannot fix that, because the chip is on the small half. Driver hardware is 16% to 20% of the finished vehicle. Take it to zero — free sensors, free compute, an ASIC that costs nothing — and an Ojai still runs about $100,000, still 3.3× to 5× a Cybercab. Every dollar of Thursday's announcement lives inside a fifth of the problem.

And the chassis saving has already been eaten. The Zeekr base vehicle is $38,000 against $75,000 for the Jaguar it replaced — 49% cheaper. After the 102.5% tariff it lands at $78,000, which is 4% more than the Jaguar. The single largest cost reduction Waymo engineered into this programme has been fully reversed by a trade measure, and the tariff alone is about 32% of the finished car — larger than the entire driver hardware package the news module was celebrating.

The announcement does not contain a car

That is the gap, and it is the axis on which Tesla is not slightly ahead but categorically ahead.

Be precise about what Waymo does and does not build. It does not build the chassis — Zeekr does, in China. What it runs is an integration plant in Mesa, Arizona, where the Driver goes into the vehicle, reported to be ramping toward tens of thousands of units a year with close to 1,000 Ojais staged there, most of them finished. That is a real industrial asset and it is not a car factory.

Tesla's Q2 2026 update lists installed Cybercab capacity at more than 125,000 units a year — over 340 a day at nameplate — on a line in continuous production since April, in Texas, tariff-free by construction. Our deployment simulation found the opposite of a production constraint: roughly 245 finished cars staged at Giga Texas in mid-July against an unsupervised fleet in the mid-teens in Austin.

Note the symmetry, because it is the fair reading: both companies have finished robotaxis parked and waiting. Waymo has around a thousand at Mesa; Tesla had 245 at Giga Texas a month ago. Neither is production-constrained today. What separates them is what the parked cars cost and what it takes to make more of them — and there Tesla owns the line, the country and the tariff schedule, while Alphabet owns the chip that reads the sensors bolted to a car somebody else builds and ships across an ocean at 102.5%.

What a million robotaxis costs

Take the fleet size out of the argument and make it a common denominator. A million cars is nobody's plan; it is the size at which the per-unit difference stops being a rounding error and becomes a capital-allocation decision. The chart above is this table:

Three readings, and the middle one is the article:

• Waymo's generational progress is enormous. $200B to $125B for the same fleet is $75B saved, and the driver-hardware column alone falls from $125B to $25B — an 80% cut. Nobody should be sniffy about that.
• It is still $95B–$105B more than Tesla, for the same million cars. That gap is 72% to 79% of Alphabet's entire trailing-twelve-month capex of $132.4B — a figure that includes every data centre it is building. The fleet itself, at $125B, is 94% of one full year of all Alphabet capital spending. Tesla's $20B–$30B is ten to fourteen months of its own guided capex of more than $25B a year, on a line it already owns.
• The chip cannot close it. Driver hardware is $25B of the $125B. Make the sensors and compute free and Waymo still spends $100B, still 3.3× to 5× Tesla, on the part of the car nobody announced anything about.

That is the shape of the problem. Waymo has spent a generation taking cost out of the layer it controls, and has done it well. The layer it does not control is now four fifths of the bill.

The cars are not a plan, either — they are live. As of 19–20 August the Ojai is open to all riders in San Francisco, Los Angeles and Phoenix, with a few hundred in commercial service, Denver, Las Vegas and San Diego queued for later in 2026, Tampa in preparation, a Hyundai Ioniq 5 platform in validation behind it, and a stated target of a million rides a week. This is not a company that needs to prove the driver works. It is a company that needs a cheaper car, and Thursday was not that.

So "Waymo is closing the cost gap" is the wrong reading of a genuinely good week for Waymo's engineering. The chip is a real step, and taken with the Marvell warrant and the inference-optimised TPU we covered yesterday, Alphabet now has custom-silicon programmes announced in the data centre and in the car inside 48 hours. But the two things that set a robotaxi's unit cost — the vehicle, and how many sensors are on it — are the two things this announcement does not address. One of them Waymo has chosen permanently. The other it has not started.

Where hardware cost stops mattering, and Waymo is ahead

The fair counterweight, and it is a heavy one.

Hardware amortises. At the $70,000 of revenue per vehicle per year our fleet piece used — our assumption, from a $165–$228 daily range — a $25,000 driver package pays for itself in about 130 days of a working car; our Tesla model carries $30,000 of revenue per deployed vehicle per quarter, roughly $330 a day, which cuts that to about 76 days. The full $125,000 Ojai takes about 652 days at the first assumption. Slower, and still finite.

What does not amortise is a person. Our fleet piece put a twelve-hour safety-driver shift at a fully loaded $22 an hour at $264 a day — more than the vehicle's entire daily revenue. One shift a day for a year is $96,360: 77% of an entire Ojai, every year, forever.

That is the ranking of costs in this business, and it puts Waymo on the right side of it. Waymo has more than 200 million miles of fully autonomous experience — its own figure, from Thursday's post — and runs paid service with nobody in the driver's seat across several US metros, while Tesla's unsupervised count has been a two-digit number for a year. Silicon that removes a person from the car is worth several times the silicon. An expensive car that works beats a cheap car that still needs a monitor, and any comparison that skips that line is measuring the cheaper vehicle rather than the operating one.

The honest summary of the two positions: Waymo has solved the driver and not the car. Tesla has solved the car and not the driver. Thursday was Waymo making its solved problem cheaper — a real step, in the right direction, on the wrong fifth of the bill. Nothing in it changes what a Cybercab costs to build or how fast Giga Texas can build one, which is what Tesla's scaling case actually rests on.

What we are not going to compare

Tesla publishes no TOPS figure for AI4 or AI4.5. What exists is a JPMorgan note, quoted in a post, from a private meeting: AI4.5 delivers "roughly 10% higher FLOPS and about 2× the memory capacity" versus AI4, and Tesla says current AI4 already supports Unsupervised FSD, with the new computer there "basically just to future-proof". That is third-hand.

Waymo's side is a front-end subtotal. Two companies, two disclosures, no shared unit, and one of the two numbers is not even the whole system. A number comparing them would have to be invented, so there isn't one in this piece. What can be said without inventing anything: a competitor published a four-figure TOPS number for sensor pre-processing in the same week Tesla was reported as saying its existing silicon is already sufficient for unsupervised driving. Two very different bets about where compute belongs, and disclosures not commensurable enough to score.

What it does to the models

Tesla — nothing, and the size of the nothing is the point. The robotaxi vertical is modelled on deployed fleet × utilisation × revenue per vehicle, with capex at 120% of segment revenue because every car is bought before it earns anything. A competitor's bill of materials lands on no line in it; there is no per-vehicle cost slider to move, because Tesla's advantage is already expressed as the thing the model does carry — capacity gated by approvals rather than by cars. Had Waymo's driver package come in at $5,000, the model would still not change. What would change it is a disclosed Waymo fleet count in metros Tesla wants, which is a demand question and not a silicon one.

Nvidia — a fourth named ASIC, in the wrong segment to matter, and Nvidia is on the customer list anyway. Our bear case names this mechanism exactly: "hyperscaler capex digests, custom ASICs take the inference workload, and merchant Ethernet claws back the fabric." Waymo joins Google's TPU line, Marvell's inference accelerator and Amazon's Trainium as another named custom-silicon programme. But the model's dcCompute vertical is built on NVL72-class racks at $3.0M, and an automotive front-end sensor processor displaces none of them — and Waymo names Nvidia as a partner in the same post, which is the opposite of a designed-out story. Where it registers is the unitPriceDrift assumption, down 1% a quarter, which the model's own note calls "what competition from custom ASICs would force". Thursday is a data point for the direction of that slider, not a reason to move it.

What to watch

1. Any cost figure from Waymo itself. There has never been one. Until there is, the $20,000, the $125,000 and the $78,000 are all somebody else's arithmetic, and this is a bounded comparison rather than a measured one.
2. The tariff line. At 102.5%, roughly half the landed chassis and about a third of the whole car is duty, and it has already cancelled a 49% chassis saving. A rate change, an exemption or US assembly moves the largest number in the stack without anyone designing anything.
3. Whether Waymo ever builds, or commissions, a cheap car. Driver hardware is a fifth of the vehicle. Everything left is the part Waymo does not make — which is why the Hyundai Ioniq 5 platform now in validation matters more to the cost line than the ASIC does. A US-built chassis removes the tariff and the ocean in one move.
4. A total-system TOPS figure from Waymo. The 1,000-plus is explicitly the front end. The ratio between front end and driving stack is the most interesting unpublished number in the post.
5. A published power budget from either company. At an analyst-estimated 1 kW the Driver eats a fifth of the Ojai's pack per day, which is a utilisation cost, not an energy bill. Neither Waymo nor Tesla publishes the number that would settle it.
6. The Mesa ramp against the Giga Texas ramp. Close to 1,000 Ojais staged against 245 Cybercabs a month ago, with a stated Waymo target of a million rides a week. Both companies have cars parked; the one that converts parked cars into paid driverless miles fastest is answering the only question that matters.

Sources and provenance. Disclosed by Waymo, in "A look under our trunk: what's in our compute" (20 August 2026) and the 6th-generation Driver post (August 2024): the 5nm ASIC and its front-end role; "over 1,000 TOPS… dedicated to front-end processing"; the "two independent engines" redundancy design; 20× compute growth in eight years; liquid-cooling integration; more than 200 million fully autonomous miles; the sensor count of 13 cameras, four lidar and six radar; and the partner list naming AMD, Micron, NVIDIA, Samsung, Sandisk, Socionext and TSMC. Waymo has published no cost figure in either post. The 45%-a-year compound rate is R40 arithmetic on the 20× claim. Press-reported: the sub-$20,000 projection for the 6th-generation driver hardware and the more-than-50% generational reduction (both attributed to Waymo but appearing only in coverage); the $25,000 figure in wider circulation; the 17-megapixel cameras; the Ojai's 93 kWh pack, 800-volt architecture and estimated 350-500 Wh/mi efficiency, and the Cybercab's reported 48 kWh pack and 165 Wh/mi — vehicle specifications as reported, not disclosures by either company, with the Waymo efficiency figure explicitly an estimate; TSMC as fabricator of this specific chip; the 5th-generation sensor counts and the 42% reduction; the Ojai opening to all riders in San Francisco, Los Angeles and Phoenix on 19-20 August with a few hundred in commercial service; the Mesa, Arizona integration plant, its reported ramp toward tens of thousands of units a year and the close-to-1,000 Ojais staged there; the Denver, Las Vegas, San Diego and Tampa expansion, the Hyundai Ioniq 5 platform in validation and the million-rides-a-week target; and the characterisation of the ASIC as matching Nvidia's autonomous-driving systems — a claim Waymo does not make. Analyst estimate: the $110,000-$125,000 range commonly quoted for a finished Ojai; the ~1 kW continuous draw of a full multi-modal Level 4 driver stack, which Waymo does not publish and which we have not independently verified; and the ~$125,000 finished Ojai and ~$200,000 Jaguar I-Pace figures, and the $75,000 per-vehicle saving, are Morgan Stanley's. Third-party estimate: the $38,000–$38,500 ex-factory Zeekr value, the 102.5% tariff and the ~$78,000 landed chassis come from one analysis of US import declarations (carnewschina, 12 August 2026); we checked its arithmetic, not its filings. R40 estimates and arithmetic, from our own published fleet math and not Tesla disclosures: the $20,000-$30,000 Cybercab build range — widened from the $30,000 in our published fleet math on the basis of the Cybercab's disclosed 48 kWh pack, two-seat interior without driver controls, and >125,000-unit line, and explicitly a case rather than a disclosure, since Tesla has never published a Cybercab build cost — plus $70,000 of revenue per vehicle per year, the $264 safety-driver shift and the 14,800-car threshold. Every derived figure here inherits their softness, including the 4.2x-6.3x and 3.4x-5.2x ratios, the driver package at 16-20% of the vehicle, the $100,000 free-hardware floor at 3.3x-5x, the $1,850M-against-$296M-$444M fleet comparison, the 130- and 652-day paybacks, the 37.5% per-vehicle cut, the 4%-dearer landed chassis, the tariff at 32% of the car, the $20B-$30B / $125B / $200B million-car totals and the $95B-$105B gap, the 80% fall in the driver-hardware column, the 20-24 kWh daily driver draw at 22-26% of the Ojai pack, the implied ranges and the 2.1x-3.0x energy ratio, the 33-50 Wh/mi attributed to the driver stack at 10-27% of the efficiency gap, the readings of those totals against Alphabet's and Tesla's capex, and $96,360 of annual safety-driver cost at 77% of an Ojai. Filed XBRL: Alphabet's trailing-twelve-month capital expenditure of $132.4B, being the four quarters to 2026 Q2 from this site's stored series. Disclosed by Tesla: Cybercab installed capacity of >125,000 units a year and capex guidance above $25B for the year, Q2 2026 shareholder update. Observer counts, not company data: the ~245 staged cars and the mid-teens unsupervised fleet. Third-hand: the AI4/AI4.5 specifications and the claim that current AI4 already supports Unsupervised FSD — a post quoting a JPMorgan note quoting a private meeting.