The Cybercab, from one clause to a production line

One sentence

Read Tesla's FY2025 annual report looking for the Cybercab and you will find it exactly once in the business description, in the future tense. The company describes launching its Robotaxi service in June 2025 as "an autonomous ride-hailing platform that harnesses our technology and vehicles," then adds the qualifier that is the whole subject of this piece: "Our Robotaxi business currently operates with Model Y vehicles but, in time, will include Cybercab, our purpose-built autonomous vehicle" ^[1].

That is the entire appearance. The vehicle that carries the company's autonomy narrative is, in the document Tesla is legally accountable for, a clause about what will happen later.

It is louder in the risk factors, and that asymmetry is informative. There the Cybercab is named as load-bearing: "our future business depends on development of our driver assistance systems and autonomous driving solutions and increasing the production of mass-market vehicles, including Cybercab, our purpose-built Robotaxi product" ^[1]. A second passage states the company "[has] plans to mass produce Cybercab" ^[1]. A filing that mentions a product once as a feature and twice as a dependency is telling you where the risk sits.

What actually carries the service

Six months after that 10-K, the Q2 2026 update reported the service running at real scale. Robotaxi "is now live in seven major metros," the company "expanded the unsupervised operation area in Austin and launched unsupervised rides in Miami, Orlando and Tampa in July," and cumulative paid Robotaxi miles reached roughly 2.5 million ^[2].

The supervision split is the detail worth holding onto. Austin, Miami, Orlando and Tampa run unsupervised; the Bay Area operates on FSD (Supervised), which the company itself qualifies — it "does not make the vehicle autonomous" ^[2]. One service, two regimes, disclosed as such.

 the Robotaxi service, as disclosed in Q2 2026 [2]
 ────────────────────────────────────────────────
 live in ................ 7 major metros
 unsupervised ........... Austin (area expanded), Miami, Orlando, Tampa
 supervised ............. Bay Area — FSD (Supervised), which the company
                          states "does not make the vehicle autonomous"
 cumulative paid miles .. ~2.5 million
 vehicle ................ Model Y  ← the Cybercab is not carrying this yet

Production began

The clause moved. In the same quarter, "Cybercab began production at Gigafactory Texas," described as "our purpose-built autonomous EV," and "engineering test drives of production Cybercabs on public roads also began during the quarter" ^[2].

What the release does not carry is a number. No Cybercab production volume, no delivery figure, no date at which it takes over from the Model Y fleet. Between January's "in time, will include" ^[1] and July's "began production" ^[2] the verb changed and the quantity did not appear. That is the honest state of the record as of this writing.

What it costs, on the operating line

Here is where the Cybercab stops being a narrative and becomes arithmetic. Tesla's Q2 2026 grew revenue hard and earned less money doing it ^[2]:

                          Q2 2025      Q2 2026      change
 ───────────────────────────────────────────────────────────
 Total revenues          $22,496M     $28,236M       +26%
 Gross profit             $3,878M      $4,751M       +23%
 Gross margin               17.2%        16.8%      -41 bp
 Operating income           $923M        $398M       -57%
 Operating margin            4.1%         1.4%     -269 bp
 Free cash flow             $146M     -$1,092M           —
 Cash & investments      $36,782M     $43,524M       +18%

Revenue up 26%, operating income down 57%, operating margin from 4.1% to 1.4% ^[2]. Operating expenses rose 47% to $4,353 million ^[2]. The company's own attribution names the causes in order: an increase in deliveries, offset by an "increase in operating expenses driven by AI and other R&D projects, SBC (including 2025 CEO Performance Award) and SG&A," lower regulatory credit revenue, and a lower vehicle average selling price ^[2].

Read that against the Cybercab and it resolves into one sentence: the autonomy programme is being paid for out of the operating line, in cash, now — free cash flow swung from +$146 million to −$1,092 million ^[2] — against revenue it has not yet produced. A $43.5 billion cash position ^[2] is what makes that a choice rather than a constraint.

The silicon underneath, and its clock

A purpose-built autonomous vehicle needs purpose-built inference silicon, and that supply chain has a timetable of its own. Samsung Foundry manufactures Tesla's AI5/AI6 self-driving inference chips at its Taylor, Texas fab on a 2nm process, with volume production expected in the second half of 2027 ^[4]. TSMC co-produces the AI5 at its Arizona fab under a dual-foundry arrangement with Samsung ^[5]. Training sits elsewhere again: the Cortex supercluster at Austin runs roughly 50,000 NVIDIA H100 GPUs, with H200s added in 2025, to train Full Self-Driving and Optimus ^[6]. These three edges are the ones this site carries on Tesla's page, each with its own source ^[7].

 train ───────────────────────────────▶ infer ──────────────▶ deploy

 Cortex, Austin            AI5 / AI6                  Cybercab
 ~50,000 NVIDIA H100       Samsung Taylor, 2nm        production began
 (+H200s in 2025) [6]      volume H2 2027 [4]         Q2 2026 [2]
                           TSMC Arizona, dual
                           foundry [5]

The dates in that diagram do not line up neatly, and the gap is the point: production of the vehicle began in mid-2026 ^[2] while volume production of the inference chip named for it is expected in the second half of 2027 ^[4]. Nothing in the cited record explains how the two are bridged — whether early Cybercabs carry prior-generation silicon, or at what rate they are built before AI5 reaches volume. That is a question the record has not answered, not one this article can resolve.

What the tape did with it

The Q2 report was published after the close on 22 July 2026. The next session answered ^[3]:

 2026-07-22   close  $374.01     report published after this close
 2026-07-23   close  $319.69     -14.52%   first post-report session
 2026-07-29   close  $298.32     -20.2% from the pre-report close
 ...
 2026-09-03   close  $376.37     +0.6% versus that same pre-report close

The move down was not ambiguous: a 14.5% single-session repricing ^[3] against a report whose revenue grew 26% and whose operating income fell 57% ^[2]. The market priced the operating line, not the top line.

The round trip is the more interesting half. Six weeks later the stock closed at $376.37 ^[3] — marginally above the $374.01 it carried into the print ^[3]. Everything the margin collapse took out, something put back. This article will not tell you what, because the cited record does not say; what it can say is what happened in that window on the record, which is that the unsupervised service expanded into three more Florida metros in July ^[2] and the Cybercab moved from a clause to a production line ^[2].

For scale, the year has still been negative: against a 2025 year-end close of $449.72, the 2026-09-03 close of $376.37 is −16.3% ^[3].

The mechanism, stated plainly

Two things are being priced against each other, and they run on different clocks.

The cost is present, cash, and measured: opex +47%, operating margin 4.1% → 1.4%, free cash flow −$1.1 billion, all in one quarter ^[2]. The return is future, unquantified, and dated only by the company's own hedged language — "in time, will include" ^[1], "plans to mass produce" ^[1], a production start with no volume attached ^[2], and an inference chip whose volume production is a year out ^[4].

A stock priced on the second while paying for the first is a stock whose price is a vote on a schedule. That is a description of the structure the cited record reveals, not a forecast of which way the vote goes. Tesla's own 10-K makes the same point in its own register, listing among the things that could go wrong the acceptance of autonomous solutions by consumers and a regulatory environment built on "many existing vehicle standards that were drafted assuming the presence of a human driver" ^[1].

What the record will answer next

Four things already named in the cited documents, worth re-reading against the next filing: whether a Cybercab production or delivery VOLUME appears anywhere ^[2]; whether operating margin stops compressing while opex growth still outruns revenue growth ^[2]; whether the unsupervised metro count moves and the Bay Area leaves FSD (Supervised) ^[2]; and whether the AI5 volume-production timing holds at second-half 2027 ^[4]. Tesla's next scheduled report is 21 October 2026 ^[7].