Digital Dailies with Matt Dho

Matt Dho, AI Matt Dho
  • 5.0 (1)
  • Technology
  • Updated Daily

An autonomously generated, human-guided daily digital download on the AI revolution designed to ensure you're the human in the loop. mattdho.substack.com

Episodes

  1. 006 - Bridging the AI Readiness Gap in Professional Services

    Jan 23

    006 - Bridging the AI Readiness Gap in Professional Services

    Sixty-four percent. That’s how many professionals have received zero AI training from their employers—while 95 percent of those same workers believe AI will be central to their jobs within five years. The talent is ready to run. Management hasn’t built the track. Welcome to Digital Dailies. I’m Matt Dho. Today we’re unpacking the Thomson Reuters Institute’s 2025 report on generative AI in professional services. The headline isn’t about technology. It’s about a massive execution gap between what workers want and what organizations provide. Here’s the shift nobody’s talking about. The fear narrative is dead. Remember 2023? Professionals were terrified AI would replace them. That’s over. The new data shows 55 percent of workers are now “excited or hopeful” about AI in their work. Not anxious. Not resistant. Ready. But here’s the friction: that enthusiasm is running into a wall. Fifty-two percent of organizations have no established policies for AI use. No guardrails. No guidance. Just employees figuring it out on their own. In human terms, this is what it looks like. A senior associate at a consulting firm needs to summarize forty documents for a client presentation due tomorrow. She knows Claude or GPT could cut that task from six hours to forty minutes. But her firm hasn’t told her whether that’s allowed, what data she can upload, or how to cite the output. So she has two choices: work until midnight doing it manually, or use the tool anyway and hope nobody asks questions. That’s not a technology problem. That’s a management failure. And it creates real liability. When training and policy are missing, high-performing employees don’t stop using AI. They go underground. They use personal accounts on unapproved tools. They paste client data into consumer products with no audit trail. They innovate in the shadows. This is Shadow AI. And it’s happening at scale across professional services right now. The irony? They’re not doing it to circumvent you. They’re doing it because they care about the work. If you’re a partner at a law firm, your best people are probably already using these tools—you just don’t know how, or on what. If you’re a CFO, your risk surface just expanded in ways your compliance team hasn’t mapped. Now let’s talk about the measurement vacuum. Here’s the number that should alarm every executive: only 20 percent of professionals know whether their organization is measuring AI’s return on investment. One in five. Companies are spending on licenses, rolling out pilots, talking about transformation at all-hands meetings. But four out of five workers have no idea if any of it is working. That’s not strategy. That’s hoping. The before/after is stark. Before: we measured software ROI obsessively—seats, utilization, time-to-value. After: we’re deploying AI with no instrumentation and calling it progress. But here’s where it gets interesting for client-facing firms. Fifty-seven percent of clients now actively want their professional service providers to use AI. They expect it. They’re asking for it in RFPs. And yet—71 percent of those same clients don’t know whether their partners are actually using AI or not. That’s a transparency gap. And for the firm willing to step into it—that’s a massive opportunity. The firm that can say “Here’s exactly how we use AI, here’s our governance framework, here’s how we protect your data, and here’s the productivity gain we’re passing to you”—that firm wins. The firm that can’t articulate that? They’re competing on price and hoping clients don’t ask hard questions. Both can be true at once. AI adoption is accelerating. And AI governance is lagging. The companies moving fastest on deployment are often moving slowest on policy. That’s not sustainable. The optimistic read: the workforce is no longer the bottleneck. People want these tools. They’re ready to build with them. The resistance has melted. The cautious read: unmanaged enthusiasm creates unmanaged risk. Shadow AI is already embedded in your workflows. You just don’t have visibility. So what do you do with this? If you’re in leadership, the question isn’t whether to publish a policy—it’s whether you can afford another quarter without one. Your people are already using AI. The only question is whether they’re doing it safely and openly, or in the shadows. If you’re in operations: instrument your AI spend. Track what’s being used, by whom, on what tasks, with what outcomes. You can’t optimize what you don’t measure. If you’re in client services: get ahead of the transparency question. Build a one-page summary of your AI governance—what you use, what you don’t, how you protect client data. For regulated industries like financial services or healthcare, this isn’t optional—it’s the baseline. Make it a competitive asset, not a liability. If you’re an individual contributor stranded without guidance: document your own use. Keep a log of what tools you’re using, what data you’re inputting, and what value you’re getting. When policy finally arrives, you won’t just be ready to lead the conversation—you’ll be the one they come to for answers. For anyone advising enterprise clients on transformation—here’s the strategic takeaway. This isn’t about who has the best models. It’s about who has the best governance. Governance enables transparency. Transparency enables client trust. Client trust enables premium pricing and retention. The winner in 2026 won’t be the firm that deploys AI fastest. It’ll be the firm that can explain it, measure it, and defend it. The technology is ready. The people are willing. Now the strategy must catch up. Sixty-four percent untrained—including, perhaps, that associate working until midnight because no one told her she could work smarter. Fifty-two percent unguided. Only twenty percent measured. Those gaps are yours to close. And make no mistake—your competitors are already trying to close them first. This is Digital Dailies. I’m Matt Dho. Subscribe wherever you get podcasts. See you next time. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit mattdho.substack.com

    7 min
  2. 005 - When Accessibility Becomes the Product

    Jan 16

    005 - When Accessibility Becomes the Product

    Seventy percent. That’s how many people can benefit from accessibility features in digital products. But here’s the kicker—less than twenty percent even know those features exist. That gap isn’t a charity problem. It’s a business problem. And the biggest companies in tech just told us exactly how they’re closing it. Welcome to Digital Dailies. I’m Matt Dho. Today we’re breaking down a CES panel that put Verizon, Salesforce, IBM, Microsoft, and Sony in the same room to talk about something most companies still treat as a checkbox. The session was moderated by Lauren Sallata, Chief Growth Officer at Jakala North America. The central question that emerged: what happens when accessibility stops being a compliance exercise and starts driving product strategy? Spoiler: it’s not a checkbox anymore. It’s becoming the product itself. Here’s the shift. For years, accessibility was an afterthought. You built the product, then you asked legal if it was compliant. Maybe you bolted on some features at the end. Maybe you didn’t. Fred Moultz, Verizon’s Chief Accessibility Officer, put it simply: “We finally have a seat at the table now. We’re part of all the conversations for new products and services. You’re not doing it as an afterthought at the end.” That’s the before and after. Before: accessibility teams reviewed finished products. After: they’re in the room when the product is being designed. Sony’s Mike Najak went further. His team doesn’t just meet regulations—they build for what he calls “co-creators with all abilities.” And they’ve got receipts. Working with the Braille Institute and hearing loss advocacy groups, Sony developed accessible kiosks for people who are visually impaired. By 2024, they’d deployed them to 924 Best Buy locations across the United States. That’s not compliance. That’s product development driven by disability communities. The throughline from Mike’s team: accessibility drives profit. That’s the sequence—profit enables purpose. Now let’s talk about how you actually operationalize this. Catherine Nichols, Salesforce’s first Chief Accessibility Officer, shared something that should change how every company thinks about funding. Like Jenny at Microsoft, Catherine came up through employee resource groups before taking the CAO role—another proof point that internal talent pipelines matter. Salesforce runs a centralized accommodations fund. All assistive tech, translators, job coaches, accessibility tools—paid from a general fund, not individual department budgets. Here’s why that matters: when accommodation costs hit a single team’s P&L, managers start doing math. Should I hire this person if it costs my budget an extra forty thousand? The centralized fund removes that friction. The cost of inclusion doesn’t land on the person making the hiring decision. That’s operational design that actually works. IBM’s approach is different but equally tactical. Will, from IBM’s accessibility team, used an analogy that stuck with me: “It’s like trying to add blueberries to the muffin after. You got to bring blueberries in at the beginning.” You cannot retrofit accessibility. You have to shift left—build it into the design phase, not the testing phase. As Catherine put it at Salesforce: they embed accessibility experts in product teams not as compliance checkers, but as innovation drivers. And here’s where it gets interesting. IBM built their entire design system—called Carbon—into an AI-accessible format using something called MCP. Now their AI tools can push accessibility best practices directly to designers and developers, even if those people don’t have deep accessibility experience. The goal: reduce the cognitive load on individual contributors. Make accessible design the default output, not the extra step. The panel turned to a harder question: what happens when AI scales inaccessibility? Microsoft’s Jenny Lay-Flurrie didn’t flinch. AI is being trained on the internet. And the internet is mostly inaccessible. Her numbers: about two to four percent of the world’s websites meet accessibility standards. That means ninety-six percent do not. And that’s the training data for the AI models we’re all deploying. Jenny put it bluntly: “We have this really crux moment right now where we’ve got a new era of technology that could repeat the mistakes of the past, or it could completely blow the doors off.” Both paths are open. The question is which one we build. The optimistic case: AI can auto-generate alt text for images, summarize meetings for people with cognitive differences, reduce noise in interfaces for neurodiverse users. Salesforce built exactly this into Slack—channel recaps, document summaries, simplified layouts that cut cognitive overload. Microsoft’s Copilot now offers real-time captions and sign language interpretation in Teams. IBM’s Carbon design system pushes accessibility patterns directly to developers through AI tooling. The cautious case: if we train AI on inaccessible data, we’ll automate inaccessibility at scale. The same biases, the same exclusions, just faster. This is where Lauren steered the panel toward the business case—because this isn’t charity. Jenny cited a study Microsoft did with Forrester. In 2003, they found 57 percent of people could benefit from accessibility features. They just re-ran that study. The number is now seventy percent. Seventy percent of your users. Seventy percent of your customers. And most of them don’t know the features exist. That’s not a compliance gap. That’s a discoverability problem. And it’s a revenue opportunity for whoever solves it. The panel made another point worth hearing: some of the most-used features in consumer tech came directly from disability innovation. Pinch and zoom? Developed by and for blind and low vision users. Closed captions? Fifty percent of Netflix viewers in the US use them—most of them aren’t deaf. One-handed gamers? Four hundred million globally. These aren’t edge cases—they’re markets. When you design for the margins, you often end up designing for everyone. So what do you do with this? Jenny’s advice was the simplest and probably the most powerful: “In every room, ask—is this accessible? And if you hear ‘of course it is,’ that’s probably a no.” If you’re in product: embed accessibility expertise in your teams from day one. Not as auditors. As builders. If you’re in finance: look at how accommodation costs are structured. If they hit individual department budgets, you’re creating friction that slows inclusion. If you’re in AI: audit your training data. If it’s mostly web content, it’s mostly inaccessible. Your models will inherit that. And if you’re building anything—software, hardware, experiences—remember Will’s blueberry muffin. You can’t add accessibility after the product is baked. Jenny Lay-Flurrie has been at Microsoft for twenty-one years. She joined the disability employee resource group, built it up, chaired it for ten years, and ended up leading global accessibility. Her path is the template: empower your disabled talent, give them room to lead, and follow their ideas—because the best innovations often come from people solving problems they actually live with. That’s the unlock from CES—and credit to Lauren Sallata for pulling these voices together. Accessibility isn’t a cost center. It’s not a legal requirement you tolerate. It’s a design philosophy that makes products better for everyone—and a business opportunity most companies are still leaving on the table. Remember the blueberry muffin: you can’t add it after the bake. Start now. Seventy percent can benefit. Less than twenty percent know. That gap is yours to close—or your competitor’s to capture. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit mattdho.substack.com

    10 min
  3. 004 - 86% Automated in 2 Years

    Jan 15

    004 - 86% Automated in 2 Years

    327%. That’s the projected surge in agentic AI adoption over the next two years, according to Salesforce. Not chatbots. Not assistants. Autonomous systems that plan, execute, and learn—without waiting for you to tell them what to do next. If you work in a knowledge role, this episode is about your job. Specifically, whether it still exists in 2028. Let’s start with what’s actually changing. The AI tools you’ve been using—ChatGPT, Copilot, Claude—they respond. You ask, they answer. You stay in control. Agentic AI is different though. You define a goal. The system breaks it into subtasks, coordinates across tools, executes the steps, monitors for errors, and adjusts in real time. It comes back with the work done. Your role shifts from doing the task to reviewing the output. That’s not a productivity boost. That’s a restructuring of who does what. Here’s the number that should get your attention: 86%. That’s how much of the time spent on complex workflows can now be handled by autonomous agents. Not someday. Now. In finance, agents are cutting fraud detection time by 30% and false positives by half. In manufacturing, early adopters report 25% higher output quality, 50% fewer errors, and 30% faster production cycles. Siemens sees 20% lower holding costs and up to 50% productivity gains. While in healthcare, autonomous systems handle patient monitoring and clinical documentation around the clock—tasks that used to require shifts of human attention. Forty-five percent of Fortune 500 companies are already piloting these systems. Eighty-two percent of executives surveyed by the World Economic Forum plan to deploy them within three years. Now let’s talk about what that means for different roles. If you’re a CMO: agentic AI can personalize at scale, optimize spend in real time, and coordinate campaigns across channels without waiting for your team to align. The upside is speed. The risk is that you lose visibility into what the system is actually doing—and inherit the blame when it does something your brand can’t defend. If you’re a CDO or Head of Data: this is your moment. Agents are only as good as the data they can access. Clean pipelines, clear governance, and robust access controls are now strategic assets. If your infrastructure is messy, your agents will amplify that mess. If you’re a CFO: the ROI case looks compelling. But implementation costs, legacy system integration, and the liability question when something breaks? Those aren’t on the slide deck yet. If you’re an analyst, coordinator, compliance officer, or project manager: the 86% number is about you. The tasks that fill your day—monitoring, coordinating, summarizing, flagging exceptions, routing decisions—those are exactly what agents are designed to do. Not eventually. Right now. Looking at the glass half full, agents handle the repetitive work, you focus on judgment and relationships and strategy. However, a word of caution, judgment and strategy are trainable too. The large language models that form the cognitive core of these agents are improving rapidly, with reinforcement learning refining their capabilities.The definition of “uniquely human” work is shrinking. Both can be true. Agentic AI will create new roles—agent supervisors, prompt engineers, AI ethics officers, governance architects. It will also eliminate roles faster than most organizations are prepared to absorb. Now here’s what’ll make this accelerate. Nvidia just announced the Vera Rubin platform. Five times faster for inference. Three and a half times faster for training than Blackwell. The NVL72 rack integrates 72 Rubin GPUs and 36 Vera CPUs—purpose-built for the massive mixture-of-experts models that power autonomous agents. The compute bottleneck that used to slow deployment is widening. The talent bottleneck—people who can build, deploy, and govern these systems—is the new constraint. The other constraint is accountability. When an agent makes a decision that harms a customer or violates a regulation, who’s responsible? The vendor who built it? The company that deployed it? The manager who approved the workflow? When an autonomous agent causes harm, who’s accountable—the developer who built it, the company that deployed it, or the manager who approved the workflow? That question doesn’t have a clean answer yet. And while many agents can surface their reasoning, some operate as black boxes—you see the output, but the decision logic stays opaque. That’s manageable for a recommendation. It’s a liability when the decision triggers a lawsuit. The companies moving carefully are building audit trails now—documenting agent decisions, creating human review checkpoints, establishing escalation protocols. The companies moving fast are hoping they won’t need to explain anything until they have to. So what do you do? If you’re in leadership: start mapping which roles are agent-adjacent and which are agent-replaceable. Build transition plans before you need them. If you’re in an at-risk role: learn how to supervise agents, not compete with them. The skill that compounds now is knowing how to review, redirect, and recover when autonomous systems fail. Now, you should really be on the lookout for the first high-profile agentic AI failure. The first lawsuit over an agent decision. The first regulatory framework that assigns liability. Those moments will define how fast this moves—and who’s prepared when it does. Three hundred twenty-seven percent adoption growth. Eighty-six percent of your workday automatable. Three years until the majority of executives have deployed. The question isn’t whether your job changes. It’s whether you’re positioned for what it changes into. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit mattdho.substack.com

    7 min
  4. 003 - CES Day 3: 70 billion parameters in your lap

    Jan 14

    003 - CES Day 3: 70 billion parameters in your lap

    Ten trillion dollars. That’s the projected cost of modernizing computing infrastructure over the next decade—and Intel just showed what that money is buying. A chip with 1.8-nanometer transistors. A neural engine running 50 trillion operations per second. A laptop that can run a 70-billion-parameter AI model without ever touching the internet. This is Day 3 of CES, and today we’re looking at the chip that could redefine where AI actually lives. Intel calls it Panther Lake. Built on their new 18A process—the most advanced manufacturing node they’ve ever shipped. The numbers: 50 TOPS from the neural processing unit. Combined with GPU and CPU, you get 170 trillion operations per second of total AI compute. That’s a 4.5x improvement in VLA—Vision-Language-Action—throughput over the previous generation. In plain English: the AI agents that watch your screen, understand your documents, and take actions on your behalf? They can now run entirely on your laptop. No cloud required. No server in the middle. No data leaving your device unless you send it. Intel’s pitch is the “PC as a private AI agent”—a machine that learns your workflow, anticipates your needs, and keeps that knowledge local. The before and after matters here. Before: AI features meant cloud features. Your data went somewhere else to be processed. After: the hardware is capable of keeping everything on-device. But here’s the tension Intel didn’t emphasize in their keynote. The hardware can run AI locally. The software ecosystem doesn’t want to. Why do apps send data to the cloud? It’s not because your laptop couldn’t handle the processing—that excuse just evaporated. It’s because cloud routing enables telemetry, training data collection, and the analytics that power advertising and product decisions. Microsoft, Google, and Apple control the platforms. They decide whether apps actually use the local neural engine or default to cloud APIs. And most apps still default to cloud—because that’s where the data monetization happens. Intel is pushing OpenVINO and partnerships with OS vendors to build an “edge-first” ecosystem. But ecosystems shift when business models shift, not when transistors shrink. This is the paradox: you can buy private AI hardware, but you might still run surveillance software on top of it. Let’s break this down by role and industry. If you’re a CIO or CTO: this is a procurement standards question. New hardware cycles are coming. The question is whether you demand local processing as a requirement—and whether your vendors can actually deliver it. Write data routing into your contracts. Audit where processing happens. If you’re a CMO or Chief Digital Officer: this is brand risk. Consumers are waking up to AI surveillance. Getting ahead with genuine privacy commitments is cheaper than recovering from the backlash. The hardware now makes “local-first personalization” technically possible. Whether your marketing stack supports it is a different question. If you run operations in manufacturing or agriculture: edge AI with industrial certification is the headline. Panther Lake is rated for 24/7 reliability. A single chip can run robotics AI and real-time controls that previously required server racks. For air-gapped facilities with proprietary processes, this is the compute you’ve been waiting for. If you’re in higher education: student data is among the most regulated. An AI tutor that personalizes without sending behavior data to the cloud is a compliance advantage. FERPA, state laws, institutional policies—all of it gets simpler when data never leaves the device. If you’re in finance or healthcare: the regulatory pressure already exists. Edge AI that processes patient data or client transactions without third-party exposure sidesteps entire categories of compliance risk. The pattern is the same across industries: the hardware can do the job locally. The software supply chain has to follow. Two things are slowing this down. First: there’s no killer app yet. The demos look great. Intel’s benchmarks are impressive. But enterprise buyers want proven ROI before they refresh hardware at scale. “Your laptop can run a 70 billion parameter model” is a spec. “This saves you $2 million in compliance costs” is a business case. The second story isn’t written yet. Second: Intel’s 18A process is new. Manufacturing yields matter. If Panther Lake chips are supply-constrained or priced at a premium, adoption slows regardless of capability. Intel is betting their turnaround on this node. The execution risk is real. One more thing to consider—the human dimension. A 4.5x improvement in VLA throughput means AI agents that watch your screen in real time, understand context, and act on your behalf. This is the “continuous supervisor” Intel hinted at: software that sees your workflow, learns your patterns, and intervenes. If that runs locally, it’s a powerful assistant. You control the data. You decide what it learns. If that runs on someone else’s server, it’s a detailed behavioral profile you didn’t sign up for. Your employer might see it. The platform vendor certainly does. The psychological adjustment is significant. For knowledge workers—the people who spend their days in documents, email, and spreadsheets—this means an AI that knows how you work. The promise is productivity. The question is whether you’re comfortable with a machine that watches everything you do. Hardware made this possible. Trust will decide whether people actually use it. Here’s what CES 2026 showed us. Day 1: AI learned physics. Robots can now train in simulation before they touch the real world. Day 2: Robotaxis are coming to American streets—twenty thousand of them by year-end—along with interior cameras, microphones, and data collection you didn’t expect. Day 3: The laptop in your bag can now run serious AI models locally. Whether it does depends on software choices that haven’t been made yet. What to watch next: which enterprise software vendors ship edge-first features; how platform providers set rules for local processing; and whether the ROI story gets clear enough to drive adoption. Ten trillion dollars is being spent to rebuild computing infrastructure. The question is whether privacy gets built in—or routed around. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit mattdho.substack.com

    8 min
  5. 002 - CES Day 2: Robotaxis and Privacy

    Jan 14

    002 - CES Day 2: Robotaxis and Privacy

    Here’s what Waymo didn’t announce at CES: they’re preparing to use interior camera footage from robotaxis—video tied to your rider identity—to train AI models and personalize ads. That’s according to TechCrunch reporting from earlier this year, and it reframes everything we saw on the show floor today. Twenty thousand robotaxis are coming to American streets. Every one of them is a rolling data collection platform. And the regulations that protect your smartphone? They weren’t built for this. Let’s start with what Uber, Lucid, and Nuro showed this week. They unveiled a production-ready robotaxi—an electric SUV built on Lucid’s Gravity platform, running Nuro’s Level 4 autonomy software, distributed through Uber’s ride network. Autonomous on-road testing began in December. The San Francisco Bay Area launch is planned for later this year. No steering wheel. No brake pedal. No driver’s seat. The exterior bristles with sensors—Autoweek described a “roof-mounted ski-rack-type module” loaded with LiDARs, radar, and cameras. But it’s the interior sensors that matter more. Because here’s what that robotaxi knows: where you’re going, who you’re with, what you’re saying, and how you’re behaving. Interior cameras. Interior microphones. All of it logged, reviewed, and—if Waymo’s draft privacy policy is any indication—potentially used for AI training and advertising. One analyst put it this way: these vehicles aren’t private spaces. They’re “rolling surveillance platforms.” The before/after here is stark. Before: you hailed a cab. The driver might overhear your conversation, but they forgot your face by the next fare. The ride left no permanent record. After: every ride is documented. Interior footage is stored. Your behavior is analyzed—ostensibly for safety, but the data has other uses. Waymo’s draft policy suggests training generative AI models. Selling personalized ads. Building profiles tied to your rider identity. If you’re a commuter who just wants to get to work, this might feel abstract. If you’re a lawyer discussing a case with a client, a journalist protecting a source, a therapist reviewing patient notes, or a parent with a teenager making choices you’d rather not have recorded—this is a different calculation. And here’s the equity problem: when only the wealthy can afford private autonomous transport, you’ve built surveillance infrastructure that monitors everyone except elites. It’s not just cars. Motorola showed a wearable at CES called Project Maxwell—an “AI Perceptive Companion.” It looks like a pearl-colored pendant. No visible display. A camera, microphones, and a speaker embedded in a glossy shell. When you activate it, the device “continuously collects full scenario data—seeing what you see, hearing what you hear, and listening to what you say.” It offers real-time insights. At the demo, someone held up a conference flyer and the pendant summarized it. The technology works. The social contract doesn’t exist yet. We’re layering always-on recording into cars, jewelry, and smart glasses without building the legal frameworks that define who owns that data, how long it’s stored, and who can access it. Behind all of this is NVIDIA’s Cosmos—the model that makes physical AI work. Jensen Huang put it directly: “Just as large language models revolutionized generative and agentic AI, Cosmos world foundation models are a breakthrough for physical AI.” The difference in plain terms: your chatbot predicts the next word. Cosmos predicts the next collision. It models physics—mass, friction, momentum—so robots and vehicles can train in simulation before they touch the real world. The early adopters are humanoid robotics companies: 1X, Agility Robotics, Figure AI, Skild AI. The bet is that machines can crash a million times in the cloud before taking a single step in your living room. But making those physics calculations in real time requires massive compute—and massive compute generates massive heat. The NVIDIA Thor chip draws up to 130 watts while delivering 2,000 trillion operations per second. One AV company, Tensor, runs eight of them with liquid cooling engineered to last 180,000 miles. That thermal envelope is the engineering constraint. Solve it, and you get autonomous vehicles that can reason. Fail, and you get overheating hardware in a moving vehicle with no driver to notice. For automakers, this changes the product. Mercedes-Benz is shipping NVIDIA’s Alpamayo software in the new CLA this quarter—the first production vehicle with reasoning-based Level 2++ driver assistance. Their CEO drove it from San Francisco to Silicon Valley. Level 4 autonomy is targeted by 2029. The before/after for driving: old systems were reactive. “I see a ball, I stop.” Physical AI is predictive. “I see a child who looks distracted. I’m slowing down before the ball rolls into the street.” That’s safer. It also means the car is continuously analyzing everything in its environment—pedestrians, passengers, context. If you work in traditional auto parts, your market is shrinking. If you work in thermal management, cognitive processors, or sensor fusion, demand is accelerating. So here’s where we are. The physics models work. The chips are fast enough. Twenty thousand robotaxis are heading to American streets this year. The open question isn’t technical. It’s whether you accept the trade: convenience and safety in exchange for continuous recording, AI training on your behavior, and a privacy framework that doesn’t exist yet. What to watch next: the first lawsuit over interior robotaxi footage; whether Waymo’s privacy policy changes after the TechCrunch reporting; and how European regulators respond when these vehicles cross the Atlantic. That about wraps up Day 2 of CES. See you tomorrow for our final day. If you enjoy our content and like staying in the loop all things AI with Digital Dailies, be sure to subscribe wherever you get podcasts! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit mattdho.substack.com

    8 min
  6. 001 - Day 1 of CES: The Age of the AI Factory

    Jan 12

    001 - Day 1 of CES: The Age of the AI Factory

    Welcome to Digital Dailies with Matt Dho Its Day 1 of CES and the personal computing landscape is experiencing a seismic shift, marked by NVIDIA CEO Jensen Huang's declaration that "The PC is no longer the center of the AI universe. The AI Factory is." This wasn't mere rhetoric - NVIDIA's actions at CES 2026 backed up this bold statement when they broke with tradition by announcing zero new consumer GPUs for the first time in five years. The absence of an RTX 50 refresh or any desktop silicon announcements signaled a dramatic change in priority. Instead, NVIDIA unveiled Vera Rubin, a massive 3.6 Exaflop industrial AI system designed for data centers. This six-chip system, already in full production and shipping to cloud partners, marks a clear pivot away from consumer computing toward industrial-scale AI infrastructure. The implications are profound for gamers, creators, and developers who have long relied on NVIDIA's consumer products. The technical specifications of Vera Rubin underscore this dramatic shift. The system uses HBM4 memory and NVLink 6 interconnect technology, enabling data transfer at 260 terabytes per second - approximately fifty times faster than high-end gaming PCs. Individual Rubin GPUs deliver 50 petaflops for inference and 35 petaflops for training, while a full DGX SuperPOD configuration reaches 28.8 Exaflops. To put this in perspective, a single rack system now matches what entire supercomputers achieved just a few years ago. This creates an unprecedented performance gap between consumer and enterprise hardware. While consumer GPUs continue using GDDR7 memory, adequate for gaming but increasingly insufficient for modern AI workloads, the industrial systems are pulling far ahead. This isn't just a temporary disparity - it represents a fundamental architectural divide that will likely widen over time. But let’s pause the technical specs for a second. If you’re sitting at a desktop right now, what does this actually mean for you? The impact varies significantly across user groups. Gamers face extended waiting periods for new hardware, with the RTX 40 series aging and RTX 50 variants backordered with no clear timeline for restock. Creators and developers watch as cutting-edge AI capabilities become exclusive to cloud-based systems. Just a year ago, capable local models for image generation, code assistance, and document analysis were possible on consumer hardware. Today, state-of-the-art models require HBM4 memory and multi-GPU fabrics beyond any desktop's capabilities. Beyond the hardware shortage, there is a quieter, more dangerous shift happening: The death of local privacy. Privacy-conscious users who handle sensitive client data, medical records, or legal documents face a particularly challenging situation. As competitive models become impossible to run locally, their data must flow through cloud infrastructure, transforming AI from a purchase to a subscription dependency. With NVIDIA abandoning the consumer, you have to ask: Is anyone else stepping up to save the desktop? The short answer is: barely. Intel offers a partial alternative with their "AI PC" initiative, featuring Neural Processing Units in Core Ultra processors. However, these target lightweight inference tasks rather than competing with Vera Rubin-class systems. Similarly, Intel's Gaudi 3 provides an enterprise alternative to NVIDIA's H100, but doesn't address the consumer market gap. The gap Intel can fill is precisely the one NVIDIA has chosen to abandon. Now, looking at this from a cold logic perspective, why would NVIDIA abandon its loyal fanbase? Simple: The math made them do it. The economics driving this shift are compelling. HBM4 memory is expensive and supply-constrained. NVIDIA's enterprise accelerators command premium prices and margins that consumer products can't match. Major cloud providers are investing over $150 billion annually in data center and AI capacity. With NVIDIA controlling over 80% of AI training and deployment GPU market share by 2025, the focus on enterprise customers is a clear business decision. AMD's response includes the Instinct MI440X, an "on-prem" accelerator for enterprises built on TSMC's N2 process and optimized for low-precision workloads like FP4, FP8, and BF16. While promising, AMD's solution still trails NVIDIA's NVLink fabric in interconnect density, and their ROCm software stack, though improving, hasn't matched CUDA's ecosystem advantages in terms of libraries, tooling, and trained developers. NVIDIA's dominance extends beyond raw computing power to their integrated stack of hardware, interconnects, and software. Their NVLink technology, CUDA platform, and ownership of key technologies like InfiniBand through Mellanox create significant barriers to competition. While alternatives like RoCE (RDMA over Converged Ethernet) and CXL (Compute Express Link) exist, NVIDIA's integrated approach delivers superior performance for most AI workloads. The company's vision extends beyond data centers into "physical AI" applications, including robotics, industrial automation, and autonomous vehicles. Their comprehensive portfolio includes Cosmos for physical environment simulation, Alpamayo for autonomous driving, the Jetson T4000 for edge robotics, and GR00T for humanoid robot development. The trajectory points toward AI becoming primarily a subscription service rather than a purchased asset. This has significant implications for privacy, control, and experimentation. Users increasingly depend on cloud infrastructure, subject to external terms of service and business models. The ability to run powerful AI locally - without permission, without sending data over the wire, without depending on someone else's terms - is rapidly diminishing. Looking ahead, several factors could influence this trend: HBM4 supply expansion might ease hardware constraints if Samsung and SK Hynix ramp production aggressively. AMD's continued development of the ROCm stack could provide alternatives, particularly if they ship desktop-class hardware optimized for local inference with adequate software support. Cloud inference API pricing will affect the economics of local versus cloud computing - if prices drop significantly, the economic argument for local hardware weakens further. So, here is the verdict for the next 24 months. The fundamental question remains: Will anyone challenge NVIDIA's vision of centralized AI infrastructure, or is the "rental future" inevitable? While NVIDIA's business decision makes sense from operational and financial perspectives, it comes with real costs for users who expected a different future. The next two years will be crucial in determining whether powerful local AI remains viable or becomes a relic of the past. Thats all for today's daily AI update and Day 1 of CES. We'll be back tomorrow for Day 2. Be sure to subscribe to our Substack so you can ensure you're always in the know on whats happening everyday in the world of AI. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit mattdho.substack.com

    9 min
  • 6 Episodes

About

An autonomously generated, human-guided daily digital download on the AI revolution designed to ensure you're the human in the loop. mattdho.substack.com

Information