It’s been several years since generative AI and large language models (LLMs) took the world by storm. LLMs surpassed earlier natural-language systems at generating text, while diffusion models enabled generating images, music, and videos.

These generative AI models work well in the digital world, but on their own, they have limited capabilities to comprehend the three-dimensional physical world and other spaces. This includes the objects occupying an area, how they relate to each other, tracking movement, and answering complex questions requiring an understanding of dimensions, distances, motion, and collisions.

Spatial intelligence is an AI capability that allows models to reason about three-dimensional space. These models can generate 3D scenes of the world and other spaces. This content can then be displayed through traditional renderers, game engines, or AR/VR systems that use spatial computing techniques. But it’s the spatial intelligence model’s ability to connect natural language with 3D models that has the most applications in robotics, manufacturing, construction, and other physical environments.   

Dr. Fei-Fei Li, often called the godmother of AI, published a manifesto on how spatial intelligence is AI’s next frontier, contrasting it with LLMs. “While current state-of-the-art AI can excel at reading, writing, research, and pattern recognition in data, these same models bear fundamental limitations when representing or interacting with the physical world,” wrote Dr. Li. “Our view of the world is holistic—not just what we’re looking at, but how everything relates spatially, what it means, and why it matters. Understanding this through imagination, reasoning, creation, and interaction—not just descriptions—is the power of spatial intelligence.”

The concept of spatial intelligence isn’t new and was described in Howard Gardner’s book, Frames of Mind, in 1983. Recent breakthroughs, including the launch of World Labs’ Marble and its $1 billion funding round, and competing approaches from Google’s Genie 3 and Nvidia Cosmos, should put spatial intelligence and world models on more R&D road maps.

What are spatial intelligence models?

It’s important to develop AI literacy and understand the terminology and concepts related to the physical world and 3D AI technologies: 

• Spatial intelligence encompasses specialized approaches such as geospatial models for mapping the physical world and building information modeling (BIM) for modeling physical structures. It also extends to generative 3D, robotics, and physical reasoning applications.
• World models are a class of neural network architectures and are currently a prominent approach to building spatial intelligence.
• Digital twins are live, virtual replicas of physical assets that combine 3D models with real-time sensor data. Spatial intelligence, an emerging capability of digital twins, adds natural-language prompting, generative scenario exploration, and physics-aware reasoning.
• Spatial computing refers to digital content anchored in and interacting with physical space, sensed and rendered in three dimensions and delivered through AR/VR and mixed-reality systems.

“Spatial intelligence models go beyond pixels to understand the 3D structure of the world—how objects are positioned, how they move, and how they interact,” says David Fattal, founder and CTO at Leia. “This enables applications like more realistic video generation, spatial computing interfaces, and AI systems that can reason about physical environments. As real-world 3D data becomes more available, these models will become foundational to the next generation of visual AI.”

Monitoring the built environment

To better understand spatial intelligence, let’s consider physical infrastructure such as bridges and buildings. The American Society of Civil Engineers estimates a $9.1 trillion investment is needed from 2024 through 2033 to achieve a state of good repair. When maintenance and monitoring lag, it can lead to major failures such as the 2022 collapse of the Fern Hollow Bridge in Pittsburgh.

Spatial intelligence and the development of digital twins may help identify issues earlier and prioritize where investments are needed. “Spatial intelligence models serve as the 4D digital blueprints for our built environment, allowing us to visualize and predict the complex interactions between aging assets and the shifting ground beneath them,” says Patrick Cozzi, chief platform officer at Bentley Systems. “By synthesizing disparate geospatial data into a living digital twin, these models provide the foresight necessary to mitigate the hidden risks of structural fatigue and subsurface instability.”

There’s a significant challenge in bridge health monitoring and transitioning from manual, infrequent structural inspections to leveraging sensors, digital twins, and spatial intelligence. Cozzi adds, “This integration of continuous field data moves beyond static documentation, empowering agencies to evolve from reactive repairs to proactive, resilient asset management that safeguards the long-term integrity of our most critical public systems.”

Avoiding collisions

Bridges are largely static, but the real world is increasingly being occupied by autonomous systems such as self-driving cars, robots, and drones. And where there are moving systems, there is a risk of collisions.

“Spatial intelligence models are AI systems that reason about the physical world by combining vision, sensor data, and contextual cues to understand space, motion, and object relationships,” says Sudeep George, CTO at iMerit. “The value of spatial intelligence models lies not just in perceiving an environment, but in enabling machines to act within it safely and in real time. That is especially important in robotics and autonomous systems, where decisions must be made in complex, multimodal, fast-changing settings.”

To see one example, this tutorial for simulating robotic environments combines Nvidia Isaac Sim, an open source robotics reference framework, with spatial intelligence in Marble from World Labs. 

Today’s collision detection systems, such as those used in autonomous vehicles, typically rely on modules for sensing, perception, planning, and control. Spatial intelligence models may offer improvements by assessing the collision risks of unidentified objects or by tracking objects that move out of sensor view. For example, Waymo’s World Model, built on Genie 3, is a simulator that generates complex weather conditions and other critical safety events.

For an out-of-this-world example, James Urquhart, field CTO and technology evangelist at Kamiwaza, has delivered several examples of spatial intelligence applications, including one for satellite collision detection and conflict analysis. Urquhart says, “Models that specialize in these types of data sets, as well as the physics and geography of the real world, enable faster and more accurate decision-making for tasks that depend on them.”

Applying spatial intelligence

Recent spatial intelligence announcements include creating 3D worlds from image or text prompts with Marble and simulating water physics, lighting, weather, and animal behavior with Genie 3. But difficulties remain in bringing spatial intelligence to physical-world use cases.

“Spatial intelligence and world models are laying the groundwork for future AI agents that will be able to interact in and with our physical world,” says Jason Corso, cofounder and chief scientist at Voxel51. “These models are significantly more challenging to develop and test, largely because the data underlying their development is complex, and it’s hard to handle all of the combinatorics involved in the physical world.”

In addition to learning the models and prototyping with them, development and data leaders need to review the data assets that will feed spatial intelligence models. “Spatial intelligence models translate location signals into a structured understanding of the real world, but they’re only as reliable as the data beneath them,” says Dan Adams, executive vice president and general manager of Enrich at Precisely. “The real unlock isn’t the model—it’s the reference layer with persistent identifiers, confidence metadata, and source lineage that lets AI reason about places, not just match strings.”

Even once applications are developed, there will be infrastructure challenges in deploying them at the edge. Ali Kayyam, principal research scientist at BrainChip, says, “The key to unlocking spatial intelligence at scale is having the low-power, event-driven hardware that can run it at the sensor in real time where it matters most.”

Where to get started

My suggestions for developers looking to get hands-on with spatial intelligence and world models:

• To try out Marble, review their API documentation and case studies, and then experiment with a developer-focused React application.
• Review the Nvidia Cosmos developer hub, documentation, and cookbook of case studies and learning paths.
• You can get an overview of Genie 3, but access is currently restricted through Project Genie, which requires a Google AI Ultra subscription.

For the past few years, the most visible corner of the AI market has been easy to caricature: OpenAI gets the consumer attention, Anthropic gets the developer love, Google gets the benefit of the doubt with increasingly capable models and a complementary product suite, and everyone else gets to explain why they’re not dead yet.

That’s unfair, of course, but not completely wrong. In AI, attention compounds and it’s leading to outsized revenue, with both OpenAI and Anthropic reportedly rushing toward trillion-dollar-sized IPOs on the backs of billions in revenue.

So it’s easy to underrate Mistral AI.

Honestly, I hadn’t thought of the Paris-based company for a year. Maybe longer. But then Brian Hall announced he’s joining Mistral as CMO, and I had an Arrested Development “Her?” moment. Hall, a longtime Microsoft exec, hired me at AWS and went on to run product marketing at Google Cloud. His move prompted curiosity because Mistral doesn’t dominate developer chatter in the United States or boast the same seemingly endless compute budgets as Anthropic or OpenAI. If the AI market is simply a race to build the biggest, most magical, most general-purpose model, Mistral isn’t the company to bet on.

But that’s the wrong question, and likely the wrong bet.

The more interesting question is when the enterprise AI market will revert to type and demand that AI deliver the same security, predictability, and control we’re used to from other IT investments. Here Mistral has a real story. As Hall notes, Mistral’s approach is to “prioritize AI for mission-critical environments that need the confidence and self-control to bet for the long term (with open weights and real sovereign capabilities).”

While this might have sounded like an overly hopeful talking point, it became real in June when the US government ordered Anthropic to suspend access for foreign nationals to its most advanced Fable 5 and Mythos 5 models. Anthropic said it would disable the models for all users because of the export-control directive. “Can this vendor be forced to turn us off?” is no longer a theoretical question.

That’s why Mistral’s quiet focus on enterprise control just might work.

The wrong race

The enterprise control story is much more compelling than the narrative I used to hear. You know, the “Europe needs its own OpenAI” schtick. There is a market for “patriotic AI,” but it’s relatively small. The far bigger market is comprised of enterprises that just want AI that works, costs less (or delivers more) than expected, and can be customized while fitting their compliance requirements.

Though the company’s initial launch page went out of its way to mention that the company was operating out of Europe and headquartered in Paris, since at least October 2023 Mistral’s product posture has centered on enterprise control. Scattered throughout its current (and past) website are words like “customize,” “fine-tune,” “open source,” and “complete control.” Mistral pitches Studio for building and running AI apps, Forge for custom model training and alignment, Vibe for agentic work, Vibe for Code for coding workflows, and Compute for training and inference infrastructure. The company talks about observability, evals, guardrails, deployment portability, and running production AI “from edge to cloud.”

In other words, it sounds less like a chatbot company and more like an infrastructure company.

That positioning becomes clearer when you look underneath the product names. Mistral AI Studio includes an AI Registry that acts as a system of record for agents, models, data sets, judges, tools, and workflows. It tracks lineage, ownership, and versioning. It enforces access controls and promotion gates before deployment. That’s boring governance plumbing (and “boring” is good in enterprise IT, as I’ve written).

Forge may be even more important. Mistral describes it as a way for enterprises to train frontier-grade models on proprietary enterprise data. Rather than training on others’ copyrighted information strewn across the web or on a mountain of Reddit posts, Forge goes well beyond retrieval-augmented generation (RAG) to not simply “read in” proprietary docs/info/etc., but rather to give an enterprise its own private OpenAI, as it were. 

That’s super interesting.

But is it different? I mean, OpenAI and Anthropic can do plenty of this, with greater scale and the benefit of leading frontier models. Both have enterprise products, cloud partnerships, evals, agents, governance tools, and varying forms of model customization. Mistral’s bet with Forge isn’t that the big labs can’t customize models. It’s that some enterprises aren’t interested in customization as a side feature bolted onto a frontier API. It is the product. OpenAI and Anthropic can build everything around Forge but not Forge itself, because the one thing they almost certainly aren’t interested in selling is independence from them.

This is where Mistral may have found a useful seam, one that allows it to ask a different set of questions. What if the best enterprise model isn’t the smartest general-purpose model? What if the best model is the one that’s small enough to run where the customer needs it, open enough to inspect and adapt, cheap enough to use broadly, and specialized enough to do the job? What if “good enough, governable, and your own” beats “slightly smarter, mostly opaque, and rented”?

This won’t matter for every use case, of course. If I’m asking AI to reason through a spreadsheet or write code, I probably want the best model I can get. But for banks, defense agencies, manufacturers, utilities, telcos, and governments, “best” is multidimensional and includes questions like latency, auditability, etc. It’s why banks, for example, still run so many workloads on premises: They want control.

What about compute?

None of this makes compute irrelevant. But it may change how compute matters.

If Mistral is trying to be a French version of OpenAI, its lack of hyperscale compute is a fatal weakness. It won’t outspend OpenAI, Oracle, Microsoft, Google, Amazon, SpaceX, or Anthropic. It probably won’t out-recruit them across every frontier research area, either. The AI market is already littered with companies that underestimated how quickly “good model” became “not good enough.”

But if Mistral is trying to become the enterprise-controlled AI layer for organizations that don’t want all intelligence to live behind someone else’s API, compute becomes a more nuanced issue. It still needs infrastructure, and Mistral seems to know it. After all, Mistral raised $830 million in debt to buy 13,800 Nvidia chips for a data center near Paris. That’s a rounding error compared to OpenAI and Anthropic, of course, but the real question is whether Mistral can turn relative compute scarcity into a virtue, like Amazon’s Leadership Principle “Frugality” on steroids. If lower compute capacity leads Mistral to deliver smaller, more efficient, and more specialized models, which in turn helps enterprises maintain more control of their data at lower cost, then less really does become more.

Mistral’s compute challenge, then, is not to try and have as much compute as OpenAI. It’s to make customers care less about raw compute scale and more about deployment flexibility, specialization, and control.

That’s a hard sell. But it’s not a dumb one.

What Mistral must prove

The bear case remains obvious. OpenAI has consumer distribution, developer mindshare, capital, and a brand that has basically become synonymous with AI. Anthropic has become the developer darling and has an unusually strong enterprise story of its own. Google has the models, the infrastructure, the data, and a bevy of complementary services. AWS, Microsoft, and Oracle have customer relationships and infrastructure.

Mistral has to prove that there’s room for another center of gravity. More specifically, it must prove three things.

First, it has to show that open-weight and controllable AI matter enough to influence buying decisions, not just conference panels. Everyone says they want control, just as most like the idea of open source. But proprietary software and cloud services still dominate the market. Mistral must make control feel like the easy button.

Second, it must prove that specialization beats generality in enough high-value markets. “Our model is almost as good” is not a strategy. “Our model is better for your bank, your government agency, or your retailer” just might be.

Third, it needs to establish a beachhead within enterprise IT before OpenAI and Anthropic become “boring” enough to satisfy the same buyers. This is the real race. The biggest AI companies are hiring enterprise sales teams, building admin controls, and cutting deals with every major cloud. Mistral’s window exists because the market is still young, but that window won’t stay open much longer.

If AI remains a model benchmark race, Mistral likely loses. But if AI keeps evolving to become grown-up enterprise infrastructure, Mistral has a real chance.

Nvidia’s new RTX Spark is one of the most interesting personal computing announcements in years. That’s because it’s not just another PC platform, but tries to redefine the role of the personal computer in the age of AI. Announced at Computex 2026, RTX Spark is Nvidia’s new platform for slim Windows laptops and compact desktops, designed to combine an Arm-based CPU, Blackwell-based RTX graphics, and a large, unified memory architecture into a single AI-first computing system.

We have all grown accustomed to a cloud-centric AI model over the past few years. We open an application, send a request over the network, and a hosted service in a distant data center provides the intelligence. ChatGPT, Grok, Gemini, and similar systems have trained the market to think of AI as something that lives elsewhere. RTX Spark proposes a different model. It asks a simple yet disruptive question: What if the model, the agent, the data, and the application could all live on your own machine? Nvidia is not just selling a faster PC. It is selling a new architectural premise.

Features, functions, and prices

On paper, RTX Spark is designed to be a highly capable local AI system. Nvidia has described the platform as combining AI acceleration and RTX graphics on a single chip for thin laptops and small desktops. Public specifications for the platform indicate configurations with up to 6,144 Blackwell GPU cores, up to a 20-core CPU, up to 1 petaflop of FP4 AI performance, and up to 128GB of unified memory. These are not ordinary PC numbers. They are clearly intended to support serious local AI workloads.

The unified memory approach is especially important. In traditional PC architecture, the CPU and GPU often use separate memory pools, which can become a bottleneck when running large models. By contrast, RTX Spark’s design is intended to make it easier for the system to host and run AI models locally. This enables Nvidia to position the machine as capable of hosting persistent AI agents, supporting local inference, and even allowing users to customize or fine-tune certain classes of language models.

Nvidia is also careful not to frame the system as only an AI box. In a smart move, the company is marketing RTX Spark for gaming, creative applications, AI development, and agentic workflows. This has been designed not as a one-trick pony, but as a capable computer first and an AI workstation second. Otherwise, it remains a niche developer experiment.

Pricing remains uncertain because Nvidia hasn’t published a universal price for every RTX Spark laptop or desktop. The platform will appear in products from different manufacturers, which means prices will vary. The best indicator comes from the related DGX Spark desktop, listed at about $4,699, though early estimates placed it between $2,999 and $3,999.

That probably gives us the right way to think about pricing for this broader category. These are unlikely to be inexpensive mainstream PCs, at least not at launch. They are more likely to arrive as premium systems aimed at developers, technical professionals, creators, and early adopters willing to pay for high-end AI capabilities on the device. Over time, that may broaden. For now, however, this looks like a new high-value, high-cost category rather than a commodity PC refresh.

What is its real purpose?

The most important thing about RTX Spark is not the chip. It is the purpose behind the chip. This machine is ultimately built to run AI agents locally, and that is a bigger deal than it may seem at first glance. An AI agent is more than a chatbot. It persists state, accesses tools, works across applications, remembers context, automates tasks, and increasingly acts as a software-based worker. Nvidia is explicitly positioning Spark systems to run personal AI agents directly on the local machine, potentially around the clock. That creates a very different computing model from what most of us use today.

There is another important layer to this story. These systems are also being positioned as platforms on which users can build and run smaller, more limited, locally tuned versions of large language model systems. Put plainly, you may be able to create your own model-based assistant that runs directly on the RTX Spark. It will not be as broadly capable as a frontier model operated by OpenAI or another hyperscaler. It is likely to be less generally capable, narrower in its expertise, and more constrained by local hardware limits. But it will be yours, it will be local, and it will respond without relying on a remote API call to a hosted AI service hundreds or thousands of miles away.

Such a shift is conceptually significant. For years, the AI industry has conditioned us to believe that serious intelligence must be centralized. RTX Spark suggests a future in which at least some intelligence becomes personal, portable, and self-contained.

Welcome to the revolution

The breakthrough here is not that local models will instantly outperform remote models. They will not. But the architecture of AI use may begin to diversify. Today, the default assumption is centralization. We assume the model, the knowledge base, and the application stack will all live in the cloud, and the user is simply a client. With systems like RTX Spark, that assumption starts to weaken. The model can run on the local machine. The agent can run on the local machine. Sensitive data can remain on the local machine. The application logic can be executed on the local machine. This changes latency, privacy, resiliency, and cost models. It also changes who controls the AI.

That does not mean the cloud goes away. Far from it. Enterprise use cases that benefit from centralized models and data will continue to exist. Businesses want the same knowledge base, business rules, database consistency, and governance model available to everyone. Centralization remains powerful because it reduces fragmentation and keeps systems aligned. Yes, a single-tenant, RTX Spark-based AI environment can be useful for certain projects, but it can also create islands of intelligence that do not easily share knowledge across teams and systems.

Possible use cases

I see the strongest potential use cases in disconnected or semi-disconnected environments. Think about physicians doing diagnostics support in privacy-sensitive contexts, field engineers collecting and interpreting data in remote areas, military and public sector users operating at the edge, or professionals who need highly private, self-contained AI assistance without relying on constant connectivity. In those scenarios, the value proposition is very strong. Having the model, data, application, and agent all on one portable system is not a limitation. It is the point.

The bigger question is whether mainstream enterprise AI will migrate in that direction. I remain skeptical that most organizations want hundreds or thousands of individually tuned, locally hosted models to replace centralized AI services. I predict that this category will complement the cloud rather than displace it. The more likely future is hybrid: centralized AI where shared knowledge and governance matter, and local AI where privacy, portability, latency, or disconnected operations matter more.

Extremely powerful large language models (LLMs) still operate as though they’re typing on a keyboard, processing workloads in a simple left-to-right fashion. But in locally-run, single-user scenarios, this sequential processing can leave graphics processing units (GPUs) and tensor processing units (TPUs) underutilized.

Google is betting that DiffusionGemma can get around this bottleneck. The new experimental open model generates text “exceptionally fast,” creating entire blocks of text simultaneously through diffusion techniques rather than through token-by-token processing. The company says this technique results in 4x faster inference compared to auto-regressive models that rely on sequential processing.

It can also save users money. Technology analyst Carmi Levy noted that existing pay-per-token monetization models “penalize the use of less than optimally efficient AI solutions.”

But DiffusionGemma “could herald a new generation of task-defined, efficient solutions that can enable expanded compute capacity without draining the operations budget,” he said.

A contrast to left-to-right processing

Built on Google’s Gemma 4 family and its Gemini Diffusion research, DiffusionGemma is a 26B mixture-of-experts (MoE) model designed to maximize text output generation.

It essentially shifts how models use hardware, giving processors a larger hunk of work each cycle so it can draft full 256-token paragraphs in sequence. This allows the model to generate text up to 4x faster on GPUs, Google claims. It activates only 3.8B parameters during inference, and, when quantized, can fit within 18GB VRAM on high-end consumer GPUs like Nvidia RTX 5090.

“It upgrades your model inference from a single, sequential typewriter to a massive printing press that stamps the entire block of text simultaneously,” Google research scientists Brendan O’Donoghue and Sebastian Flennerhag wrote in a blog post.

AI image generators begin with pure, random ‘visual noise’ and iteratively refine that into a finalized picture (what’s known as ‘diffusion’); DiffusionGemma applies this same process to text. It does not generate tokens in order, but begins with a “canvas of random placeholder tokens” that it processes in multiple passes, identifying the context tokens it feels are most relevant and using those to refine the rest.

The model has the ability to self-correct, using confidence scoring to re-evaluate tokens in the next pass. “The model iteratively refines its own output, allowing it to evaluate the entire text block at once to fix mistakes in real-time,” O’Donoghue and Flennerhag explained.

DiffusionGemma also has bidirectional attention, they wrote. “Generating 256 tokens in parallel with each forward pass allows every token to attend to all others.” This can be particularly helpful in domains that are non-linear in nature, such as mathematical graphs, code infilling, and in-line editing, they said.

DiffusionGemma is optimized across Nvidia’s hardware stack, making it compatible with consumer setups as well as with high-performance enterprise systems like Hopper and Blackwell.

Because it is released under the Apache 2.0 license, developers can freely use, modify, distribute, and commercialize the software using their preferred tools. It can be run on GPUs or in the cloud through Google Cloud Model Garden or Nvidia NIM, and is available on Hugging Face, GitHub, and vLLM, with support for the open-source library llama.cpp coming soon.

Key use cases

The model is particularly useful in local workflows that are “speed critical,” such as generation of non-linear text structures, and unlocks what Google calls “new patterns of model behavior” like multimodal understanding and generating and rendering code in near real-time.

Levy explained, “DiffusionGemma is particularly well suited for interactive coding and editing where its efficiency allows rapid processing and iterations,” noting that its ability to fit within 18GB of VRAM and its deployability on commonly available local GPUs can potentially benefit customer service-related workloads that lean heavily on real-time interaction and local processing.

“DiffusionGemma also incorporates a thinking mode that is especially adept at problem solving,” he said. For instance, the model was fine-tuned to play Sudoku, a typically challenging task for autoregressive models because each token depends on future tokens. This “rather handily” illustrates the model’s capability to solve more complex problems, Levy noted.

Limitations

Google freely admits that DiffusionGemma is geared to specific workflows, and there are “key trade-offs.”

The model is engineered for small batch size inferencing and low-latency, high-speed generation low-to-medium batch sizes on a “single capable accelerator.”

In high-QPS cloud serving environments, (where infrastructure is designed to handle tens or hundreds of thousands of requests per second with ultra-low latency), DiffusionGemma’s parallel coding “offers diminishing returns,” and can even result in higher serving costs, Google conceded. In addition, its overall output quality is lower than that of standard Gemma 4, which is built for apps demanding maximum quality.

However, Levy noted that while DiffusionGemma “can be less precise than other models in certain workloads,” subsequent refinement cycles could overcome this limitation.

While Google isn’t sharing runtime costs, it’s clear that this is an efficiency play, he added. “When deployed across the kinds of workloads that would optimally benefit from its architecture, DiffusionGemma seems to have the potential to reduce processing overhead and related costs,” he said.

This article originally appeared on InfoWorld.

Extremely powerful large language models (LLMs) still operate as though they’re typing on a keyboard, processing workloads in a simple left-to-right fashion. But in locally-run, single-user scenarios, this sequential processing can leave graphics processing units (GPUs) and tensor processing units (TPUs) underutilized.

Google is betting that DiffusionGemma can get around this bottleneck. The new experimental open model generates text “exceptionally fast,” creating entire blocks of text simultaneously through diffusion techniques rather than through token-by-token processing. The company says this technique results in 4x faster inference compared to auto-regressive models that rely on sequential processing.

It can also save users money. Technology analyst Carmi Levy noted that existing pay-per-token monetization models “penalize the use of less than optimally efficient AI solutions.”

But DiffusionGemma “could herald a new generation of task-defined, efficient solutions that can enable expanded compute capacity without draining the operations budget,” he said.

A contrast to left-to-right processing

Built on Google’s Gemma 4 family and its Gemini Diffusion research, DiffusionGemma is a 26B mixture-of-experts (MoE) model designed to maximize text output generation.

It essentially shifts how models use hardware, giving processors a larger hunk of work each cycle so it can draft full 256-token paragraphs in sequence. This allows the model to generate text up to 4x faster on GPUs, Google claims. It activates only 3.8B parameters during inference, and, when quantized, can fit within 18GB VRAM on high-end consumer GPUs like Nvidia RTX 5090.

“It upgrades your model inference from a single, sequential typewriter to a massive printing press that stamps the entire block of text simultaneously,” Google research scientists Brendan O’Donoghue and Sebastian Flennerhag wrote in a blog post.

AI image generators begin with pure, random ‘visual noise’ and iteratively refine that into a finalized picture (what’s known as ‘diffusion’); DiffusionGemma applies this same process to text. It does not generate tokens in order, but begins with a “canvas of random placeholder tokens” that it processes in multiple passes, identifying the context tokens it feels are most relevant and using those to refine the rest.

The model has the ability to self-correct, using confidence scoring to re-evaluate tokens in the next pass. “The model iteratively refines its own output, allowing it to evaluate the entire text block at once to fix mistakes in real-time,” O’Donoghue and Flennerhag explained.

DiffusionGemma also has bidirectional attention, they wrote. “Generating 256 tokens in parallel with each forward pass allows every token to attend to all others.” This can be particularly helpful in domains that are non-linear in nature, such as mathematical graphs, code infilling, and in-line editing, they said.

DiffusionGemma is optimized across Nvidia’s hardware stack, making it compatible with consumer setups as well as with high-performance enterprise systems like Hopper and Blackwell.

Because it is released under the Apache 2.0 license, developers can freely use, modify, distribute, and commercialize the software using their preferred tools. It can be run on GPUs or in the cloud through Google Cloud Model Garden or Nvidia NIM, and is available on Hugging Face, GitHub, and vLLM, with support for the open-source library llama.cpp coming soon.

Key use cases

The model is particularly useful in local workflows that are “speed critical,” such as generation of non-linear text structures, and unlocks what Google calls “new patterns of model behavior” like multimodal understanding and generating and rendering code in near real-time.

Levy explained, “DiffusionGemma is particularly well suited for interactive coding and editing where its efficiency allows rapid processing and iterations,” noting that its ability to fit within 18GB of VRAM and its deployability on commonly available local GPUs can potentially benefit customer service-related workloads that lean heavily on real-time interaction and local processing.

“DiffusionGemma also incorporates a thinking mode that is especially adept at problem solving,” he said. For instance, the model was fine-tuned to play Sudoku, a typically challenging task for autoregressive models because each token depends on future tokens. This “rather handily” illustrates the model’s capability to solve more complex problems, Levy noted.

Limitations

Google freely admits that DiffusionGemma is geared to specific workflows, and there are “key trade-offs.”

The model is engineered for small batch size inferencing and low-latency, high-speed generation low-to-medium batch sizes on a “single capable accelerator.”

In high-QPS cloud serving environments, (where infrastructure is designed to handle tens or hundreds of thousands of requests per second with ultra-low latency), DiffusionGemma’s parallel coding “offers diminishing returns,” and can even result in higher serving costs, Google conceded. In addition, its overall output quality is lower than that of standard Gemma 4, which is built for apps demanding maximum quality.

However, Levy noted that while DiffusionGemma “can be less precise than other models in certain workloads,” subsequent refinement cycles could overcome this limitation.

While Google isn’t sharing runtime costs, it’s clear that this is an efficiency play, he added. “When deployed across the kinds of workloads that would optimally benefit from its architecture, DiffusionGemma seems to have the potential to reduce processing overhead and related costs,” he said.

Several trends are now converging that threaten to pit tech companies against the general public. 

Miniaturization has finally enabled companies to build AI glasses that look and function like normal glasses, but with microphones and cameras. People are increasingly talking to AI, rather than typing. And multimodal input, especially video, is on the rise. 

Put all of these trends together and you get a nascent industry pushing toward all-day, everyday AI glasses with cameras — and a worried public already pushing back at  the idea.

Let’s look at how we got here.

Meta started it with a surprise hit: its second-generation Ray-Ban Meta glasses, which later gained multimodal AI capability. Its Meta Ray-Ban Display glasses add one in-lens screen — but both versions of the glasses have cameras. (The company is working on a third generation that will probably ship next year.)

Google provides the AI and software platform through Android XR and Gemini, partnering with hardware makers to put its AI on other companies’ glasses. At Google I/O last month, Google unveiled frames from Gentle Monster and Warby Parker running Android XR with Gemini AI; they’re scheduled to launch this fall. Google is working on two types of AI glasses, one with screens and one that is audio-focused. Both types have cameras, though. 

Samsung is working to launch AI-powered smart glasses, too, code-named “Jinju.” The company offered up details at Google I/O alongside Google. The glasses feature a 12-megapixel camera with autofocus; run on Android XR with Gemini AI; are co-designed with Gentle Monster and Warby Parker; and are slated to launch in July at the Samsung Unpacked event. 

(As with Meta and Google, Samsung is working on AI glasses with and without screens, but both of its models have cameras.)

Tech giant Apple is also on the glasses train, based on reporting from anonymous insiders. Codenamed N50, the Apple glasses could have two cameras, one for pictures and videos, the other for multimodal AI input and hand-gesture control. (Apple is also working on a pendant and next-gen AirPods, both of which have cameras.) 

There’s Amazon, which is reportedly developing a new line of consumer AI glasses with a camera after its earlier, camera-less Echo Frames and Carrera Smart Glasses lines failed. (My guess is the problem was Alexa, not the lack of cameras.) Although its Echo Frames have been effectively discontinued — displayed as sold out online — the company is already testing AI glasses with cameras for enterprise use on hundreds of US-based Amazon drivers. 

Huawei in April launched its AI Glasses for the Chinese market — the lightweight glasses sport a dual-engine AI architecture and integration with its HarmonyOS ecosystem. It’s joined there by Xiaomi’s AI Smart Glasses, which are powered by the company’s HyperOS ecosystem and have cameras for photos and videos and for for reading QR codes. 

Beyond those well-known firms, other companies are making daily-wear AI glasses with cameras in them, including XREAL, Rokid, TCL, Solos, and Brilliant Labs.  

A minority of other companies is focused on glasses without cameras, including Even Realities (G1 and G2); MIRA (MIRA glasses); Dymesty (Dymesty AI glasses); Lucyd (Lucyd Lyte); and Huawei (Eyewear 2).

Get the picture?

Clearly, by the end of the year, the market will be flooded with all manner of AI glasses designed for everywhere, everyday wear. They can use prescription lenses or serve as sunglasses — and most of them will have cameras built in for photos, videos and multimodal AI. 

There’s just one problem: The public hates AI glasses with cameras.

Return of the ‘Glassholes’?

As we learned from Google Glass, a lot of people feel uncomfortable with a camera pointed at them while they’re talking to someone. And that backlash is back with the current generation of AI glasses. 

Because Meta is the market leader in the US, its Ray-Ban Meta glasses have borne the brunt of early disaffection. 

Texas Attorney General Ken Paxton recently launched a formal investigation into Meta’s AI glasses, calling them “a privacy nightmare for Texans,” claiming the devices “can easily invade personal privacy by collecting biometric data and recording Texans without their knowledge or consent.” 

Paxton also claimed the LED light on the glasses, which is designed to alert others that the camera is taking pictures or videos, can be easily defeated. In fact, some modders-for-hire charge up to $100 to physically destroy the LED and TikTok videos describe how to disable or cover the light. 

The pushback is happening elsewhere. Philadelphia courts banned smart Meta AI glasses with recording features from city courthouses and a petition is circulating to ban them from New York City bars and restaurants. MSC Cruise Line banned smart glasses in all public areas. And restaurants, gyms, and workplaces have begun banning smart glasses because of the camera. 

Uncertainty drives some of the concern. People don’t know whether they’re being recorded, and if they are, they don’t know who will see the video. It turns out, those  suspicions might be warranted. 

In February, Swedish publications Svenska Dagbladet and Göteborgs-Posten published an investigation that found Meta contractors in Kenya were reviewing footage from Ray-Ban Meta smart glasses — including “bank details, sex and naked people who seem unaware they are being recorded.” 

The New York Times published an internal Meta memo in February describing plans to add facial recognition (“Name Tag”) to Ray-Ban Meta glasses. The memo said the “political tumult in the United States would distract critics from the feature’s release.”

Then earlier this month, WIRED discovered dormant facial-recognition code called “NameTag” hidden inside Meta’s AI companion app. The code would let Ray-Ban Meta glasses identify strangers by face, a feature Meta publicly claimed “does not exist.” Meta quietly erased the code with an update one day after the exposé was published.

A coalition of civil society organizations wrote Congress to demand that Meta abandon its Name Tag facial recognition plans, calling it a “creepy and unacceptable escalation of surveillance.” The letter warned the technology could be adopted by law enforcement to surveil immigrants, people of color, and nonviolent protesters.

Finally, a range of reports involving AI glasses with cameras in them has emerged in recent months involving secret recording, harassment and extortion. 

The coming conflict over face cams

On one hand, all the biggest consumer electronics companies are either shipping AI glasses with cameras in them or planning to do so — and many smaller companies are looking to do the same. The industry expects AI glasses with cameras to go totally mainstream. 

On the other hand, a growing public, legal and legislative backlash has erupted in opposition to AI glasses with cameras in them. 

One possible outcome is that the public disdain for the cameras will fade, overwhelmed by widespread enthusiasm for the benefits they offer. A new social norm might emerge that mirrors the broad acceptance of everybody having cameras in their phones and pointing them in random directions. 

Another possibility is that companies will be forced by consumer disdain and legal action to abandon cameras in glasses and focus instead on AI glasses that can’t take pictures or use video for multimodal AI input. 

Either way, the war is surely coming. 

Anyone can build an app now. But nobody seems to care.

Well, not nobody. VCs keep funding startups that add AI to, well, everything. But users aren’t buying the massive influx of new apps. In a chart shared by Jen Zhu Scott based on the new National Bureau of Economic Research’s working paper “Writing Code vs. Shipping Code,” iOS app releases have exploded since the advent of agentic AI. That would perhaps be cause for celebration had app reviews not declined during this same period, and apps with significant usage have stayed essentially flat.

In other words, more apps but almost nobody new showing up to use them.

For those of us that grew up in open source, it’s a familiar problem. The greater the abundance of code, the greater the need to help would-be customers navigate it through marketing (including branding), sales, etc. AI is creating so much noise, in terms of new code, new products, etc., that the real work has shifted to taste-making.

Getting more but not using more

I’ve been saying for a while that developer productivity isn’t about producing more code faster. Or at least it shouldn’t be. Productivity is about producing well-architected, secure, maintainable code that solves a problem someone actually has.

That’s a very different thing. It’s not something AI can fix; at least, not yet.

Charity Majors put it more bluntly in a post I wrote in 2024: “Writing code is the easiest part of software engineering,” she said. The harder parts are figuring out what to build, integrating it into a larger system, validating that it works, maintaining it over time, and getting humans to trust it enough to use it.

Turns out the harder parts are really hard. Mert Demirer, Leon Musolff, and Liyuan Yang tracked more than 100,000 GitHub developers alongside their AI usage telemetry. Autocomplete, interactive agents, and autonomous agents each ramped raw coding activity, with cumulative effects on commits of 40%, 140%, and 180%. That sounds great until you look how the gains attenuate the closer you get to actual users. For example, that 180% jump in commits became roughly 50% more projects and just 30% more actual releases. The report’s authors call this the weak-link problem: The strong link (writing code) got much stronger, while the weak links (everything else humans have to do) didn’t. The estimated elasticity of substitution between AI and human effort is 0.25, which is economist for “these complement each other, but they don’t replace each other.”

When the authors checked four major app marketplaces, they found a bump in new apps but no increase in total usage. In other words, we’re getting better at creating things with AI’s help; apparently we’re not good at turning that into user interest. User attention (and budget) is finite. The hard part is to figure out how to get users to care enough to pay (with their time or their money).

Learning to love marketing

So what’s actually scarce in a world of near-infinite software? Not code, for sure. No, what’s scarce is attention, trust, and a reason to switch.

That means the durable advantages in software increasingly live in the parts of the business developers often undervalue, like a recognizable brand (something that Red Hat figured out early on in open source) or a channel they already use. Or it could be in areas that developers appreciate but don’t pay for, like good documentation or a welcoming community.

This isn’t a new idea; it’s just newly unavoidable. We’re watching something similar inside enterprises, where AI adoption is wildly uneven, not because the technology has no value, but because the organizational plumbing around it often hasn’t been built. This is why a new kind of speed is important. I once argued that speed was the killer app, and that was mostly true. Today the more interesting speed is how quickly you can earn trust, fit into a workflow, answer objections, and get adopted.

Years ago I wrote that RethinkDB was dead and MongoDB wasn’t what killed it. RethinkDB was at the time, by many technical measures, a better database, built around “correctness, simplicity, and consistency.” It still lost—and badly. Its founder’s own postmortem was unsparing: They had picked a brutal market and tuned the product to the wrong definition of good. Being technically right turned out to have almost nothing to do with getting adopted.

With AI we’re generating a thousand smaller “RethinkDBs,” only faster and with nicer landing pages generated by the same model that wrote the code. These aren’t going to win, any more than RethinkDB was able to unseat MongoDB. It’s not just about the tech, but rather about making that tech fit within a user’s or enterprise’s world and making it easy to adopt. AI makes this harder, not easier.

It’s a cliché that developers don’t like marketing. It’s also false in my inexperience. What developers don’t like is traditional marketing. During my time at MongoDB and now at Oracle, my developer relations teams have focused on offering developers deep, hands-on enablement, and the response has been fantastic. Is a technical tutorial marketing? Of course it is. So is a forward-deployed engineer working side by side with an enterprise’s engineers to help them effectively use your tech. Just because it’s not a 30-second Super Bowl ad doesn’t mean it’s not marketing.

Years ago the Dilbert cartoon captured developers’ disdain for marketing. It was funny then, and it’s funny now, but it has never been true.

Years ago Matt Klein, creator of the open source Envoy project, once talked me through all the non-development work that goes into making an open source project thrive. As he reflected, “If you look at what I did in 2016 and early 2017 to [introduce] and grow the project, it was not technical.” So what was it, if not core engineering? “It was all leadership, public relations, marketing, documentation, etc., and I did it all myself and I nearly killed myself [doing it].” As he summed it up for me, it was a “f—ing lot of work,” and much (most?) of it wasn’t about code. It was about helping users appreciate the value of that code.

TL;DR? The boring but essential go-to-market grind has always been the real job. It’s what will separate winners from losers in AI.

The boring work wins

None of this is an argument against AI coding tools. Developers should absolutely use them. I use them constantly.

But AI can’t compensate for the hard work that goes into making a product successful in the market. In a world drowning in AI-generated sameness, taste becomes a competitive advantage. For me, “taste” translates into knowing what not to build, or what not to publish. It’s also all about knowing how to market one’s product, which might include ads but definitely needs to incorporate technical training that helps developers make sense of your code.

In short, there’s no Field of Dreams “build it and they will come.” There’s just a lot of hard, human work to help real people find and use your code.

Agentic AI has moved from conference hype to a budget line item. This is where the conversation gets more interesting and more uncomfortable. Unlike traditional AI systems that respond to a single prompt, classify a document, recommend an action, or generate a summary, agentic AI systems are designed to pursue goals. They plan, call tools, inspect results, retry failed steps, consult memory, hand off tasks to other agents, and sometimes critique their own work before producing an answer or taking an action.

That extra autonomy is the value proposition. It also introduces the cost problem.

A single chatbot interaction may consume a few thousand tokens. A useful agentic workflow can consume hundreds of thousands or millions of tokens per day because it does more than answer a question. It decomposes the problem, retrieves context, reasons through options, invokes APIs, checks the output, and often runs multiple passes before reaching a result. Therefore, the economics need to be understood at the level of “agent instances,” not just model calls.

For the estimates below, I am using a blended token cost of $3 dollars per million tokens. This is not intended to reflect a single vendor’s list price. It is a blended planning figure that assumes a mix of input and output tokens, reasoning steps, retrieval-augmented generation, summarization, tool calls, memory updates, and occasional use of larger context windows. Some enterprises will pay less through volume discounts or by routing work to smaller models. Others will pay more by using premium models, long-context prompts, web browsing, large document ingestion, and repeated reasoning loops.

The basic formula is straightforward. If an agent consumes 2 million tokens per day, it consumes 730 million tokens per year. At $3 per million tokens, that single agent costs about $2,190 per year in token burn. That number sounds surprisingly low until you multiply it by the number of agents, workflows, and users, plus the surrounding infrastructure required to run these systems safely.

What an agent really costs

In the model used here, the annual token-only cost per agent ranges from about $1,095 to $3,833, depending on the use case.

• A lightweight HR recruiting or onboarding agent that consumes one million tokens per day costs about $1,095 per year.
• A more demanding software engineering agent that consumes 3.5 million tokens per day costs about $3,833 per year.
• Customer support agents are roughly $2,190 per year.
• Legal contract agents cost about $2,409 per year.
• Security triage agents run about $2,738 per year.
• Research agents are around $3,066 per year.

These figures are useful but incomplete. They include only LLM token consumption and exclude orchestration platforms, vector databases, observability, model evaluation, security controls, workflow monitoring, human review, enterprise application integration, data pipelines, audit logging, prompt management, and the engineers needed to build and maintain the systems. In real deployments, I would expect the all-in operating cost to be two to five times the raw token cost. For regulated or mission-critical environments, the multiplier can be even higher.

This is where many agentic AI business cases become less clear. The model call may be inexpensive, but the system around the model is not. An agent that can update a customer relationship management system, approve a refund, generate a purchase order, or recommend a security containment action needs guardrails, permissions, logging, rollback mechanisms, and human escalation paths. These are not optional features. They are the difference between a demo and an enterprise system.

Use cases that make economic sense

Customer support is one of the more obvious use cases. A typical support automation deployment may use eight agents: an intake classifier, a knowledge retrieval agent, a response drafting agent, an escalation agent, a quality review agent, a CRM update agent, a sentiment detection agent, and an analytics agent. At two million tokens per agent per day, each agent costs about $2,190 per year in token burn, bringing the annual total to roughly $17,520. If that system deflects even a modest number of tickets or improves agent productivity, the economics can be attractive.

Sales development is another practical example. A five-agent system for account research, lead enrichment, email personalization, CRM updates, and follow-up scheduling may consume 1.2 million tokens per agent per day. That results in an annual cost of about $1,314 per agent, or $6,570 for the full agent team. This can be compelling if it improves pipeline quality, but it can also be wasteful if agents generate low-quality outreach at scale. The cost of brand damage is not measured in tokens.

Software engineering is more expensive but potentially more valuable. A 12-agent system covering requirements analysis, architecture, code generation, testing, review, security checks, documentation, CI debugging, refactoring, release notes, dependency analysis, and hot-fix support may consume 3.5 million tokens per agent per day. That works out to about $3,833 per agent annually or roughly $45,990 for the full system. Compared with engineering salaries, the token cost is small. The real question is whether the system reliably improves throughput without increasing defects, security vulnerabilities, or maintenance complexity.

Security operations also fit the agentic model because the work is repetitive, time-sensitive, and context-intensive. A 10-agent security triage system could include agents for alert triage, log analysis, threat intelligence, endpoint investigation, network investigation, incident summarization, ticketing, compliance evidence, escalation, and post-mortem. At 2.5 million tokens per agent per day, the annual token cost is about $2,738 per agent or $27,375 for the system. This is easy to justify if it reduces alert fatigue and accelerates response, but risky if the agents hallucinate causality or bury critical signals in confident summaries.

Finance, legal, healthcare administration, market research, HR, and supply chain are also viable:

• A six-agent finance close system may cost about $9,855 per year in token burn.
• A four-agent legal contract review system may cost about $9,636.
• A seven-agent healthcare administrative workflow may cost about $13,797.
• Competitive intelligence using six agents may cost $18,396.
• HR recruiting and onboarding using five agents may cost $5,475.
• Supply chain planning and exception management using eight agents may cost $21,024.

Across all 10 example use cases I’ve mentioned in this section, the model assumes 71 agents and a total annual token burn of about $175,638.

The comparison with traditional AI

The economics of agentic AI should always be compared with simpler approaches. Traditional AI, workflow automation, rule engines, robotic process automation, and non-agentic LLM calls are often cheaper, easier to govern, and more predictable. Agentic AI is usually overkill for tasks like classification, extraction, summarization, routing, or drafting within a narrow context. A deterministic workflow with a single model call can do the job at a fraction of the cost and risk.

Agentic systems make sense when the process requires judgment across multiple steps, dynamic planning, tool use, exception handling, and adaptation to incomplete information. They are valuable when the path to the answer cannot be fully scripted in advance. They are less valuable when enterprises use them as a trendy substitute for basic automation.

The best architecture is usually hybrid. Use traditional automation where the process is stable. Use non-agentic AI where the task is bounded. Use agentic AI only where autonomy creates measurable leverage. That means fewer agents, tighter scopes, explicit budgets, model routing, token monitoring, and human checkpoints for high-impact decisions.

The financial mistake many organizations will make is to treat agents as digital employees with near-zero marginal cost. They are not. They are probabilistic software components that consume tokens, trigger tools, create operational dependencies, and require supervision. The token bill may be manageable. The governance bill may not be.

Agentic AI can absolutely be worth the money. In many cases, the annual token burn for a useful agent team is less than the loaded cost of a single employee. But that is not the same as saying it is cheap. Companies must measure agent cost per completed business outcome, not cost per prompt or cost per model call. In the end, the question is not how much an agent costs. Ask yourself the right question: Does the autonomy an agent provides outweigh the complexity it introduces?

Forget “Florida Man.” Want to hear a California Man story? 

Here goes. 

A California man rolled up to a yoga studio in San Francisco’s Marina District in a self-driving Waymo car, walked into the studio, grabbed an armful of yoga shorts, got back in the Waymo and took off. 

Six months later, police still haven’t found him, according to a story this week in The San Francisco Chronicle. Since the rider’s credit card information didn’t lead to an arrest, we can assume the perp used a stolen phone’s Waymo account and financial information to hail the ride. And by the time police requested interior video of the man’s face, Waymo had already deleted it. 

This is a “California Man” story in part because of the association of Waymo with the city of San Francisco. Soon that association will be obsolete. (In fact, while Waymo is headquartered in San Francisco and is more visible there, Arizona got Waymos two years before San Francisco did.) 

At the moment, Waymos are publicly available to riders in 11 US cities — San Francisco, Phoenix, Los Angeles, Austin, Atlanta, Miami, Orlando, Dallas, Houston, San Antonio, Nashville.

Before long, riders will be able to enjoy robot car rides from Waymo in Las Vegas, San Diego, Washington DC, Denver, Detroit, Baltimore, Boston, Charlotte, Chicago, Minneapolis, New Orleans, New York, Philadelphia, Pittsburgh, Sacramento, Seattle, St. Louis, Tampa, London, and Tokyo.

Historically, Waymo has been taking a bath on rides. 

(I’m not talking about recent stories where Waymo cars have driven onto flooded roads. On May 12, Waymo issued a voluntary recall of 3,791 cars after a software defect allowed an autonomous vehicle in April to drive into a flooded, impassable roadway in San Antonio and be swept into a creek. A week after the recall, the company paused all freeway rides and suspended service in Atlanta, Austin, Dallas, Houston, San Antonio, and Nashville because of construction-zone navigation issues.)

To date, self-driving ride-hailing services like Waymo are a loss-leader business. Waymo is secretive about its costs, but independent estimates suggest that a $20 ride for the rider may be a $50-$100 ride for Waymo, when you factor in all costs. 

But help is on the way. 

Cars that drive themselves don’t pay for themselves

A big part of why Waymo rides have been so costly is that the car is a retrofitted Jaguar I-PACE. It’s true that Waymo got a deal on the roughly $70,000 car (a steal at $50,000 per vehicle because the company bought thousands of them). 

But then Waymo had to bolt on all kinds of costly sensors and electronics to make them self-driving, including a roof-mounted lidar assembly of five units, 29 cameras all around the car, six radar units, a custom Waymo-designed AI inference compute platform, and the wiring harness and power distribution system. 

Estimates for the total cost per car for Waymo are in the $120,000 to $200,000 range. 

Another problem is that the Jaguar I-PACE is notorious for a lack of reliability, especially involving its batteries and its longevity. Jaguar stopped making I-PACE cars two years ago. 

Finally, Waymo can’t do what regular car owners do and sell the car to recover some of the initial investment. Nobody wants an electric car with a depleted battery covered in electronics and sensors that can’t be used. 

The good news is that we learned this week that used Waymo batteries will be repurposed as backup storage for power grids in California and Texas. 

Say “Oh, Hi!” to the Ojai

The combination of growth and the end of manufacturing for the Jaguar I-PACE means that Waymo’s next platform is right on time. The car is called the Ojai, named after the unaffordable artsy hippie mecca located 80 miles northwest of Los Angeles. 

Waymo announced last week that the company will soon open Ojai cars to free rides for select riders in San Francisco, Los Angeles, and Phoenix. (See this Redditor’s drone photos  of the current fleet of Ojai cars in Mesa, AZ.)

The Ojai is a purpose-built electric minivan made by Zeekr, an arm of China-based Geely Automobile Holdings. It’s got doors that slide open like an elevator door, more legroom than the I-PACE, three screens for passengers, Braille instructions, grab bars, a flat floor with low-step height for easier entry, charge ports, cupholders, more cargo space, better batteries, faster EV charging, and easier cleaning and maintenance than the I-PACE cars. 

It’s much cheaper for Waymo to buy — and much cheaper and faster to integrate Waymo electronics. 

The Ojai gets Waymo’s sixth-generation Driver as a factory-co-engineered system with just 23 sensors (13 cameras, four lidar, six radar). While that’s far fewer sensors, they’re much more capable than the older generations of Waymo systems. One of my favorite details of the Ojai sensor package is that each car will have 10 sensor wipers with heaters and fluid sprayers specifically designed for snow, rain, and adverse weather. They’re like tiny, high-tech windshield wipers, but for the glass in front of the sensors. Ojai cars will likely do far better in rain and snow conditions. 

Weirdly, the Ojai still has needless controls, like a steering wheel and gas and brake pedals. And the only reason for those is that the US Congress is asleep at the wheel. US. federal motor vehicle safety standards require steering wheels and pedals for street-legal vehicles, and neither NHTSA nor Congress has granted a permanent exemption for purpose-built driverless vehicles. 

Despite the vestigial controls, the difference between the two cars is that the Jaguar was an old-school car designed for drivers, while the Ojai is the new concept for cars, one built for riders. And for riders, the Ojai is better in every way that matters. 

The Chinese factor

There’s only one factor keeping Ojai cars from replacing the full Jaguar I-PACE: They’re  made in China. 

Waymo is getting around the 100% tariff imposed by the Trump Administration by “location-laundering” the build. Zeekr completes the Ojai shells in Gothenburg, Sweden, and because the “substantial transformation” occurs within the EU, the vehicles are classified as EU-origin products. The stripped-down gliders arrive with no modems, ECUs, or autonomy software, and Waymo installs all connected technology at its Mesa facility, which satisfies the Commerce Department’s 2027 and 2030 rules prohibiting Chinese-linked vehicle electronics. 

Some lawmakers are using Waymo as a case study for general anxieties about Chinese technology infiltration into American infrastructure. The other problem is protectionism. If un-tariffed Chinese cars were allowed into the US market, the US car industry would likely be decimated by the competition. Chinese carmakers like BYD enjoy a 25% material cost advantage over Western carmakers. They would enter $5,000 to $10,000 cheaper than comparable U.S. offerings, according to some estimates. 

So, Washington is jittery about Chinese-made cars. 

I drove a BYD rental car in the UK last month. And I can tell you, they’re great cars and very enjoyable to drive. (My only complaint was that the steering wheel was on the wrong side.)

Instead of Waymo taking a risky bet on Ojai cars, they’re instead expanding with Hyundai Ioniq 5 EVs, which are produced locally and will be retrofitted with Waymo’s sixth-generation Driver at that Mesa facility. This is a massive deal in which Hyundai will supply Waymo with 50,000 cars by 2028. 

Waymo hasn’t disclosed plans for Ojai cars, but it’s unlikely to even come close to the number of Hyundai cars it is on the hook for. 

(The company also has around 100 Zeekr cars, but plans to expand that fleet to a few thousand.)

The right solution for Waymo’s next few years would be all Ojai cars with no steering wheel or pedals. The Ojai is purpose-built for autonomous car ride sharing, affordable, and capable in all weather. But that’s not going to happen because of Congressional inaction, China panic, and protectionism in Washington. 

Instead, Waymo’s future is to use too many cars from the past, by which I mean much or most of its fleet will be driverless cars retrofitted from cars that prioritize the driver, rather than the passengers. And reports suggest that the price of Hyundai cars will be comparable to the overpriced Jaguars. 

I’m sure the Hyundai Ioniq 5 will be nice. But an all-Ojai fleet would have been the better future for Waymo. Instead of the right car for the job, Waymo is stuck with an expensive, less comfortable, less capable car than the Ojai.