Zoox Turns Its Driverless Shuttle Into a Paid Las Vegas Service
The purpose-built robotaxi offers a distinctive cabin and app-based ride, but its first commercial footprint remains narrow.
Zoox will begin charging passengers for robotaxi rides in Las Vegas on Monday, August 10. The commercial launch turns a free public demonstration into a transport product, but it does not make Zoox a generally available ride-hailing network.
The service is unusual because the vehicle was designed without a driver’s position. There is no steering wheel, pedal box or conventional front seat. Two rows of passengers face one another inside a compact electric carriage, and the vehicle can travel in either direction without turning around.
For a passenger, the basic workflow is familiar. The Zoox app shows whether service is available, accepts a destination and summons a vehicle. Riders can pair a phone over Bluetooth, use wireless charging at each seat and contact live support through the app.
What changes next week is the transaction. Reuters reports that Zoox will use a base fare plus time and distance, with pricing comparable to the “comfort” tier of conventional ride-hailing services. That tier is commonly priced above a standard ride. Zoox has not published a complete fare table, so the cost of a specific journey will be visible only when a rider requests it.
Most robotaxi services begin with a production car and add sensors, computers and automated-driving software. This preserves familiar controls and makes it possible for a human to take over during development.
Zoox took the opposite approach. Its vehicle has symmetrical ends, sliding doors and inward-facing seats. Removing the driver’s area gives the cabin a social, shuttle-like layout and lets every seat be treated as a passenger seat.
That design is the product’s clearest distinction, but it also created a regulatory problem. U.S. vehicle standards contain requirements written around human-operated cars, including controls and equipment associated with a driver’s forward view. Zoox could not simply treat those features as irrelevant and begin commercial service.
On July 31, the National Highway Traffic Safety Administration granted a temporary exemption covering the purpose-built vehicle. The decision runs for two years and permits commercial deployment of up to 2,500 vehicles annually. It includes enhanced oversight and allows the agency to update operating authorizations as the technology develops.
The exemption is permission to deploy eligible vehicles under federal safety rules. It is not a declaration that Zoox may operate anywhere. State permits, local operating conditions, pickup access and the company’s own service map still determine where a passenger can request a ride.
Zoox currently lists Las Vegas, San Francisco, Austin, Miami, Atlanta and Los Angeles on its support site. That list combines different stages of testing and passenger access. Paid rides are beginning only in Las Vegas, and the company has not given a date for charging in other cities.
Even within Las Vegas, the service is not equivalent to entering any address in a mature ride-hailing app. Zoox has built its public operation around selected destinations near the Strip and has expanded that network over time. Its own Las Vegas page tells users to download the app and check current availability rather than promising citywide coverage.
The practical product is therefore a managed route network. A visitor whose hotel, restaurant or attraction is supported can request a distinctive driverless ride. Someone outside the operating area still needs another form of transport.
An independent field report published in May illustrates the operational challenge. Road to Autonomy recorded waits of 48 and 67 minutes during two Las Vegas requests. It also described a limited set of destinations and inconsistent wireless charging. Those observations predate the paid launch and represent two rides, not a comprehensive service audit. They are still a useful reminder that ride quality includes vehicle supply, pickup time and routing, not only autonomous driving.
Charging fares may improve demand management because fewer people will order a vehicle only for the novelty. It can also expose a harder test. Paying customers compare wait time, destination coverage and reliability with ordinary taxis and ride-hailing apps.
The commercial start proves that Zoox has moved beyond free promotional rides in one market. It also gives the company real pricing and demand data for a vehicle with no manual controls.
It does not yet prove that the service can scale economically. Purpose-built vehicles require dedicated manufacturing, maintenance, charging, cleaning and remote support. A small operating area can simplify those systems, but expansion multiplies the number of edge cases and facilities the company must manage.
Zoox also remains responsible for software updates and incident reporting. The federal exemption is paired with additional oversight rather than a permanent waiver from scrutiny. The company issued a recent software recall related to how vehicles responded near smoke at emergency scenes, showing why rare conditions matter even after routine rides work.
For early Las Vegas users, the decision is simpler. Check the app, compare the quoted fare and wait time, and confirm that both endpoints sit inside the service area. The product is now a paid ride rather than a technology preview. Its most important limitations are visible before the door opens: where it goes, how long it takes to arrive and what the app asks the passenger to pay.
Anthropic Is Building a Chip Team, Not a Chip Factory
The company confirmed plans for custom silicon, but its existing AWS, Google, Nvidia and AMD relationships remain central.
Anthropic confirmed on August 5 that it is building an internal team to design custom chips for Claude. The move gives the AI company a new route to improve performance and control costs, but it is much earlier than the phrase “Anthropic chip” may suggest.
The most concrete evidence is a new Silicon Engineer vacancy on Anthropic’s careers site. The role covers front-end design, verification, physical design, analog systems, foundry technology, packaging and chip testing. It asks for candidates who have personally helped a semiconductor reach production and who can support the first chip when it arrives.
That is a serious hiring brief. It is not a product announcement. Anthropic has not named a processor, disclosed specifications, identified a manufacturing partner or provided a delivery date.
Reuters reported that the company wants engineers across hardware and software to co-design chips and AI models. Anthropic described custom silicon as another part of a multi-chip strategy. It said technology from Amazon Web Services, Google, Nvidia and AMD will remain in the mix.
Large AI models perform enormous numbers of matrix calculations. General-purpose graphics processors handle this work well, but a chip designed around one company’s models can remove features it does not need and accelerate operations it uses constantly.
The benefit is not only raw speed. Hardware and model teams can choose memory formats, data movement, numerical precision and networking together. For inference, which is the process of producing an answer from a trained model, small efficiency improvements become important when repeated across millions of requests.
Anthropic’s job description makes that co-design explicit. The new engineer would work with inference, performance, software-kernel and infrastructure teams. The listing also says the team will decide what to build, buy or license, and when to rely on partners.
That last point matters. Designing a chip does not require Anthropic to own a fabrication plant. Most specialist chip companies create the architecture and design files, then hire a foundry to manufacture the silicon and other partners to package and test it. Even a design developed internally may incorporate licensed processor interfaces, memory controllers and networking technology.
The hard part is making the whole system work at scale. A useful AI accelerator needs access to high-bandwidth memory, fast links between servers, stable software tools and enough power and cooling. A competitive chip without that surrounding system is not a substitute for a mature computing platform.
Anthropic is already committed to outside infrastructure on a very large scale. In April, it said it would spend more than $100 billion on AWS technology over ten years and secure up to five gigawatts of capacity. That agreement spans several generations of Amazon’s Trainium chips.
The company also expanded work with Google and Broadcom. Anthropic says Claude runs across AWS Trainium, Google TPUs and Nvidia GPUs so workloads can be matched to different hardware. AMD is now included in its stated multi-chip strategy as well.
Those commitments make the new team easier to interpret. Custom silicon can give Anthropic negotiating leverage, supply diversity and a design tailored to selected Claude workloads. It does not need to replace every external accelerator to be useful. An internal chip might first target a narrow job such as inference, model training or data processing.
Business Insider reported that the company is offering between $320,000 and $485,000 in annual salary for the Silicon Engineer role. The listing describes a small team with broad responsibilities, another indication that the program is at a formative stage rather than preparing a finished product for launch.
The company has not said whether it will lead the complete design or share responsibility with an application-specific integrated circuit partner. It has not confirmed a foundry, manufacturing process or whether the first design will be used only internally. Reuters notes that an advanced AI chip can cost roughly $500 million to design.
There is also no evidence yet that Claude runs faster or more cheaply on Anthropic-designed hardware. Those outcomes can be measured only after a design is completed, manufactured, integrated with software and operated under real workloads.
The news is still significant because it changes Anthropic’s role in the supply chain. The company is moving from choosing and optimizing other companies’ processors toward owning at least part of the silicon roadmap. For customers, however, the immediate reality is unchanged. Claude continues to depend on a diverse collection of cloud providers and chip makers, while Anthropic begins the long process of adding one more option of its own.
How 6G Beamforming Can Focus a Signal at a Point
Large antenna arrays can use wavefront curvature to target both direction and distance, but the hardware trade-offs are still unsettled.
Most cellular antennas aim a radio beam in a direction. A phone somewhere along that path can receive the signal, much as an object in a flashlight beam catches light. A proposed technique for 6G is more selective. It can concentrate radio energy around a particular point, defined by both direction and distance.
This is called near-field beamforming, or beam focusing. The name does not mean the phone must be almost touching the antenna. With the very large antenna arrays being studied for 6G, the relevant near field can extend for tens or even hundreds of metres.
A paper published in Scientific Reports on August 5 compares several ways to build such a system. Its most useful contribution is not a declaration of one winner. It shows why the best architecture changes with distance, bandwidth and the cost of the radio hardware.
Radio waves spread outward in curved fronts. Far enough from an antenna array, however, a small section of that curve looks nearly flat. Current beamforming systems usually exploit this approximation. They adjust the phase, or timing position, of the signal at each antenna element so the waves reinforce one another in a chosen direction.
The approximation becomes less accurate as the antenna panel grows, the wavelength shrinks or the receiver moves closer. The boundary is commonly described by the Rayleigh distance. It grows with the square of the array's physical aperture and decreases with wavelength. In other words, a larger array or higher radio frequency pushes the boundary farther away.
Within that boundary, each part of the array is measurably farther from the receiver than another. The incoming or outgoing wavefront must be treated as spherical rather than planar. Those path-length differences encode two coordinates. The system can infer and control angle, as before, but also distance.
To focus a transmission, the array calculates the path from every antenna element to the target point. It then applies a delay or phase adjustment so the many copies of the signal arrive together there. Away from that focal point, they align less neatly. The IEEE Communications Society compares the effect to a controllable convex lens.
This extra distance dimension could let a base station serve two devices that sit along nearly the same line but at different ranges. Research published this year on volumetric beam focusing goes further, exploring ways to concentrate energy inside a compact three-dimensional region rather than across an entire direction.
The most flexible design would give every antenna element its own complete radio-frequency chain, including conversion and digital processing. That is fully digital beamforming. It offers fine control, but thousands of chains would consume substantial power and add cost.
Hybrid beamforming divides the work. An analog layer uses phase shifters or true-time-delay components to coordinate many antenna elements. A smaller number of digital radio chains then separates users and data streams. This retains much of the spatial control without duplicating expensive electronics behind every element.
The new Scientific Reports study compares fully digital processing with several hybrid layouts. In a fully connected layout, every radio chain can influence every antenna. A sub-connected design assigns each chain to a smaller group. Delay-phase precoding combines phase shifters with time delays, which matters because a phase setting that focuses one frequency may misalign another across a wide channel.
The paper also evaluates reconfigurable intelligent surfaces. These are electronically controlled panels that redirect radio waves rather than generating the main signal themselves. In the study's simulations, the surface-assisted method had the highest modeled energy efficiency. Delay-phase processing was the more consistently adaptable option across the tested distances and bandwidths.
Those results describe a model, not a field deployment. The reported rankings depend on the power assumptions, channel conditions and selected configurations. The article does not show a commercial 6G base station serving moving phones.
A focused link requires an accurate description of the channel between the array and the device. In the far field, a network can search a codebook of directions. Near-field training must search direction and distance, increasing measurement and signaling overhead.
Movement makes that harder. A narrow focal region is useful only if the network can estimate position, update the beam and calibrate a large array quickly enough. Reflections and blockage add more paths to the calculation.
Wide bandwidth creates another problem called beam split. Different frequencies can focus at different locations when the hardware applies the same phase pattern across the band. True-time-delay components can reduce the effect, but they add complexity and power loss of their own.
Standardization is also incomplete. The IEEE overview notes that established 5G channel models largely assume far-field propagation, although 3GPP work is extending models for higher bands and near-field behavior. Real measurements, shared channel data and hardware prototypes are still needed to determine where the theoretical gain survives practical constraints.
The important shift is conceptual. A large array is not merely a louder or narrower antenna. Once wavefront curvature becomes useful information, the network can treat depth as a communications resource. The engineering question for 6G is how much hardware and computation that extra dimension is worth.