On Tesla’s Q3 2026 earnings call on 21 October, one of the top retail shareholder questions, backed by about 1.1 million shares according to Electrek, was simple: when does unsupervised FSD arrive in customer cars? Ashok Elluswamy answered with a date, around Q2 2027, and a definition. In his words, as quoted by The Verge, the software for Cybercab and Model Y “is the same”, it will ship to customers as part of the V15 release, and “the only additional requirement is to make sure that, for example, the customer Model Y can be part of the geofences and the fleet management that Tesla has for it to be fully autonomous.”

My conclusion up front: Tesla has quietly changed the unit of acceptance. “Unsupervised FSD” used to be described as a property of a car, meaning your car drives itself wherever you take it. As described on this call, it is a property of a fleet operating inside a zone, with your car allowed to join. That is a sounder spec than the old promise. It is also a different product from the one owners were sold, and it should be judged on fleet-and-zone evidence, not on the software version number.

What I concluded

First, “same software” does not mean “same product”. The sentence that matters is the one about geofences and fleet management. A robotaxi’s operational design domain (ODD) is not defined by the network weights alone. It is defined by the mapped and validated zone, the remote assistance and recovery behind it, the maintenance and cleaning cadence, and the operator who carries the liability. Saying the owner’s car must sit inside Tesla’s geofence and fleet management is a statement that the safety case lives there, not in the car.

Second, this is the right call as engineering, and it is narrower than the marketing. A personal car that drives itself anywhere has an ODD equal to wherever the owner goes. That is very hard to validate, insure or permit, which is why Tesla’s own Robotaxi service runs inside bounded zones. Electrek describes Tesla’s current Robotaxi service as six US metros, mostly small slices of each city, largely off highways, and still with a safety driver in the Bay Area. By framing the Q2 2027 step as owner cars working inside those zones, Tesla is choosing the only version that can actually be accepted. The cost is that it is not what Electrek says Musk has sold to customers since 2016, a car that drives you anywhere while you sleep. Electrek also quotes Musk telling investors in July 2025 that unsupervised FSD would be available “for unsupervised personal use by the end of this year in certain geographies.”

Third, the business is already pointing this way. CNBC quotes Musk saying “the vast majority of the Cybercab stuff will be essentially Tesla-owned and transport-as-a-service.” Tesla’s own Q3 outlook says deliveries will depend partly on “allocation decisions between sale to customers and use for our owned and operated fleet.” Read together, the owner-car step looks like a capacity valve for a fleet Tesla runs, not a consumer feature released to owners.

Constraints

The binding constraint is fleet heterogeneity. A Tesla-owned fleet in a zone is one configuration with one maintenance regime. Owner cars arrive with different hardware generations, different sensor-cleaning histories, tyre and brake wear Tesla did not schedule, and owners who may need the car back at short notice. Electrek notes that HW3 owners have already been told their computers cannot run unsupervised FSD. Admitting owner cars means the safety case has to include an admission test per vehicle, and that test is a product in its own right.

The second constraint is the denominator. Tesla’s Q3 update says the Robotaxi service reached “nearly two million cumulative unsupervised miles” as of the end of the quarter, 30 September. Tesla’s Robotaxi safety page states 3.1 million miles driven autonomously as of 18 October 2026, for 22 January through 18 October, with no safety monitor in the vehicle. From 30 September to 18 October is 18 days. Subtracting the filing’s nearly 2 million from that 3.1 million leaves about 1.1 million miles, and neither document explains the jump. That unexplained increment has to be closed before strangers’ cars join the fleet.

The third constraint is regulatory and legal, and it is per jurisdiction. The Verge points out that driverless ride-hailing needs specific local and state permits, that insurance may be hard to obtain, and that in most places it is still not legal for individuals to run their own self-driving cars as driverless taxis. In California, Electrek reports that state rules still require a safety driver in every Tesla Robotaxi. A Q2 2027 date for owner cars is therefore a date for specific zones in specific states, whatever the software is ready to do.

The demand side is real, and that is what makes the reframing sensitive. Tesla reports FSD (Supervised) attach above 75% of new North American deliveries, up nearly twenty percentage points from the previous quarter, and 1.74 million active subscriptions, up from 1.48 million in Q2. CNBC puts the US price at $99 a month. Most of those buyers live and drive outside any Robotaxi zone.

The options, and what I traded away

My reconstruction of the choice set, not a description of anyone’s internal reasoning.

Evidence

The strongest evidence is the wording. The VP answering the shareholder question defined the deliverable as joining Tesla’s geofences and fleet management, not as a feature unlock. Tesla’s filed update supports the fleet reading: Cybercab rides began on 3 September in Austin, Austin hours are now 6am to midnight, cumulative unsupervised Robotaxi miles grew eightfold quarter on quarter, and the outlook explicitly weighs selling cars against keeping them for Tesla’s own fleet. CNBC reports Texas DMV counts of 362 Model Y Robotaxi vehicles and 449 Cybercabs authorised for driverless operation. The mileage gap is the finding. The filing’s nearly two million cumulative unsupervised miles are as of 30 September; the safety page’s 3.1 million miles with no safety monitor are as of 18 October. About 1.1 million miles in 18 days is left unexplained.

Where this stops being true

Segment: this is about unsupervised operation. For supervised FSD the car is still the right unit of acceptance, because the driver is the fallback, and the 75% attach rate is evidence about that product, not the unsupervised one. Region: in a jurisdiction that licenses individual owners to run driverless vehicles, the fleet framing becomes a choice rather than a necessity, and the argument weakens. Fleet size: with a few hundred cars per city, a Tesla-owned fleet is easier to keep uniform than a mixed one; at tens of thousands, owner capacity may be the only way to grow fast enough, and the admission test becomes the main engineering problem. Model and hardware generation: if V15 needs newer hardware, the share of owner cars that can even apply may be small, which makes the Q2 2027 milestone smaller than it sounds. Company size: my read is that few companies other than Tesla have both a large owner fleet and a robotaxi operation, so the lesson for most OEMs is the acceptance-unit point, not the fleet-sharing model. And if Tesla ships per-vehicle unsupervised operation outside any geofence and it holds up, my reading was too conservative.

What I would do differently

State the acceptance unit in the first sentence of any autonomy promise. “Your car, anywhere” and “your car, inside our zone, run by us” are different products, and customers should hear which one they are buying.

Publish the admission criteria with the date: minimum hardware, sensor and camera cleanliness checks, maintenance state, software version, and how a car is removed from the fleet. That turns “around Q2 2027” into a checklist owners and regulators can track.

Reconcile the safety-page mileage with the filing’s cumulative unsupervised miles before admitting a single owner car. About 1.1 million miles separate the nearly two million as of 30 September from the 3.1 million as of 18 October, and that gap is still unexplained.

Keep improving supervised FSD as its own product, with its own success metrics, and stop letting it stand in for progress on unsupervised operation. The two are now different businesses with different acceptance tests.

中文版

问题

10 月 21 日的特斯拉 2026 年三季度财报电话会上，据 Electrek 报道，散户股东提问中排名靠前的一条（约 110 万股支持）很直接：客户自己的车什么时候能用上无监督 FSD？Ashok Elluswamy 给出了日期，大约 2027 年二季度，同时也给出了定义。据 The Verge 引述，他说 Cybercab 和 Model Y 的软件“是同一套”，会随 V15 版本推送给客户，“唯一的额外要求是，比如说，确保客户的 Model Y 能纳入特斯拉的地理围栏和车队管理，才能实现完全自动驾驶。”

我的结论先放在这里：特斯拉悄悄换掉了验收单位。“无监督 FSD”过去被描述成一辆车的属性：你的车，你开到哪里它就能自己开到哪里。按这次电话会的说法，它变成了一个车队在某个运营区内的属性，你的车获准加入。这个规格比原来的承诺更站得住，但它和当初卖给车主的不是同一个产品，应该用车队加运营区的证据来评判，而不是看软件版本号。

我的结论

第一，“同一套软件”不等于“同一个产品”。关键是那句地理围栏和车队管理。Robotaxi 的运行设计域（ODD）不只由模型权重决定，它由经过采图和验证的运营区、背后的远程协助与救援、保养清洁的节奏，以及承担责任的运营方共同决定。要求车主的车必须纳入特斯拉的地理围栏和车队管理，等于承认安全论证在那里，不在车上。

第二，作为工程决策这是对的，但它比营销说法窄。一辆能随处自动驾驶的私家车，ODD 就是车主会去的所有地方，这样的范围很难验证、投保或拿到许可，所以特斯拉自己的 Robotaxi 也只在有边界的运营区内运行。Electrek 描述特斯拉目前的 Robotaxi 覆盖美国六个都市区，多数只是城市中的一小片，基本不上高速，湾区仍配安全员。把 2027 年二季度这一步定义为“私家车在这些运营区里干活”，特斯拉选的是唯一能真正通过验收的版本。代价是，它不再是 Electrek 所说的马斯克自 2016 年以来卖给客户的东西：一辆能在你睡觉时载你去任何地方的车。Electrek 还引述马斯克 2025 年 7 月对投资者说，无监督 FSD 将在当年底前“在部分地区开放给个人无监督使用”。

第三，业务本身已经指向这里。CNBC 引述马斯克的话：“Cybercab 的绝大部分基本会是特斯拉自有、以出行即服务的方式运营。”特斯拉三季度展望写道，交付量部分取决于“在卖给客户与留作自有运营车队之间的分配决策”。两者合起来看，私家车加入这一步更像是给特斯拉自营车队加的一个运力阀门，而不是推送给车主的消费级功能。

约束条件

最硬的约束是车队的异质性。特斯拉自有车队在一个运营区里是同一种配置、同一套保养制度。私家车则带着不同代际的硬件、不同的传感器清洁状况、特斯拉没排过的轮胎和刹车磨损，车主也可能临时要用车。Electrek 提到，HW3 车主已被告知其计算平台无法运行无监督 FSD。接纳私家车，意味着安全论证里必须包括逐车准入检测，而这套检测本身就是一个产品。

第二个约束是分母。特斯拉三季度更新写的是，截至季末（9 月 30 日），Robotaxi 服务累计无监督里程“接近 200 万英里”。特斯拉 Robotaxi 安全页写明，截至 2026 年 10 月 18 日，无安全员的自动驾驶里程为 310 万英里，统计区间是 1 月 22 日至 10 月 18 日。从 9 月 30 日到 10 月 18 日是 18 天。用 310 万减去接近 200 万，差额大约 110 万英里，两份文件都没有解释这次跳跃。这个没有解释的增量，得先补上，才能把别人的车放进车队。

第三个约束是监管和法律，而且按辖区分别生效。The Verge 指出，无人驾驶网约车需要具体的地方和州级许可，保险可能很难买到，而且在大多数地方，个人用自己的自动驾驶车辆做无人出租车仍不合法。Electrek 报道，加州的规定仍要求特斯拉 Robotaxi 每辆车配安全员。所以私家车的 2027 年二季度，只能是特定州、特定运营区的日期，与软件是否就绪无关。

需求端是真实的，这也正是这次改口敏感的原因。特斯拉报告，北美新车交付中 FSD（监督版）选装率超过 75%，环比提升近 20 个百分点；活跃订阅 174 万，二季度为 148 万。CNBC 给出的美国价格是每月 99 美元。这些买家大多数住在、开在任何 Robotaxi 运营区之外。

取舍选项

以下是我对选项集的重构，不是任何人的内部推理。

证据

最强的证据是措辞本身：回答股东提问的副总裁，把交付物定义为纳入特斯拉的地理围栏和车队管理，而不是解锁一项功能。特斯拉提交给 SEC 的报告支持车队这个解读：Cybercab 于 9 月 3 日在奥斯汀开始载客，奥斯汀运营时间现为早 6 点到午夜，累计无监督 Robotaxi 里程环比增长八倍，展望部分明确在“卖车”和“留给自营车队”之间权衡。CNBC 报道，得州机动车管理局数据显示，获准无人驾驶运营的有 362 辆 Model Y Robotaxi 和 449 辆 Cybercab。里程缺口本身就是发现。财报里接近 200 万英里的累计无监督里程截至 9 月 30 日；安全页上 310 万英里的无安全员里程截至 10 月 18 日。18 天里大约 110 万英里没有得到解释。

这个结论的边界

细分场景：本文讨论的是无监督运营。对监督版 FSD 来说，车仍然是正确的验收单位，因为驾驶员是兜底，75% 的选装率是那个产品的证据，不是无监督产品的证据。地区：如果某个辖区给个人车主发无人驾驶运营许可，车队框架就从必然变成选择，论证随之减弱。车队规模：每城几百辆时，自有车队比混合车队更容易保持一致；到几万辆时，私家车运力可能是唯一足够快的扩张方式，准入检测就成了主要工程问题。模型与硬件代际：如果 V15 需要更新的硬件，能申请加入的私家车比例可能很小，2027 年二季度这个节点就比听起来小。公司体量：我的看法是，除了特斯拉，很少有公司同时拥有庞大的私家车保有量和 Robotaxi 运营，所以对多数主机厂来说，可借鉴的是“验收单位”这一点，而不是车队共享模式。最后，如果特斯拉真的在任何地理围栏之外实现了逐车无监督运营并且站得住，那就是我的判断过于保守。

我会怎么改

在任何自动驾驶承诺的第一句话里写明验收单位。“你的车，随处可用”和“你的车，在我们的运营区内，由我们运营”是两个产品，客户应该听清自己买的是哪一个。

和日期一起公布准入标准：最低硬件、传感器与摄像头清洁检测、保养状态、软件版本，以及车辆被移出车队的条件。这样“大约 2027 年二季度”就变成一张车主和监管都能跟踪的清单。

在接纳第一辆私家车之前，先把安全页的里程和财报里的累计无监督里程对上。9 月 30 日的接近 200 万英里和 10 月 18 日的 310 万英里之间大约差 110 万英里，这个缺口仍然没有解释。

把监督版 FSD 当作独立产品持续改进，给它单独的成功指标，不要再让它代替无监督运营的进展。两者现在是验收标准不同的两门生意。