Colossus Compute Wars, Tesla's Robotaxi Reality Check, and the Wearable Biotech Wave — This Week in Tech

The AI Infrastructure Arms Race Just Got Very Personal

If you want to understand where generative AI is really heading, stop watching the model benchmarks and start watching the data centers. This week, Bloomberg reported that Elon Musk's SpaceX ran into serious trouble trying to develop and run its Grok AI models inside its own Colossus facility in Memphis, Tennessee. The problem? Latency issues when connecting Colossus 1 with two other sites more than ten miles away, compounded by aging network infrastructure. So SpaceX did what any Silicon Valley giant would do: it started renting out its supercomputer to its competitors.

Anthropic agreed to pay $1.25 billion per month through May 2029 for access to Colossus I and Colossus II. That is $15 billion annually — nearly doubling SpaceX's total 2025 revenue of $18.7 billion. Google followed with its own deal, reportedly paying $920 million per month. Both agreements include a 90-day termination clause, which makes perfect sense when you're dealing with an industry that moves faster than most contracts can keep up with.

Inside SpaceX's S-1 filing, the numbers are staggering. The company spent $12.7 billion in capital expenditures on AI in 2025 — roughly 61 percent of its total capex. In the first quarter of 2026 alone, AI spending hit $7.7 billion, dwarfing the $1 billion SpaceX spent on its space division. The AI division lost $6.3 billion on $3.2 billion in revenue in 2025, and lost another $2.5 billion in Q1 2026. It is a spectacular burn rate by any standard, which makes the decision to rent out capacity to Anthropic and Google not just sensible but necessary. The IPO itself opened at $135 per share, valuing SpaceX at roughly $1.77 trillion before trading pushed it higher, briefly making Elon Musk the world's first trillionaire.

This is not just a SpaceX story — it is a window into how the entire AI market operates right now. Training frontier models requires clusters of tens of thousands of GPUs running in sync for weeks or months. A single training run for a model comparable to GPT-4o or Claude can cost tens of millions of dollars in compute alone. The hardware — NVIDIA H100 and H200 GPUs, custom ASICs, high-speed networking fabric — is in severe shortage, and the companies that already own large data centers have enormous leverage. SpaceX built Colossus at hyperscale, and now it is monetizing that investment the only way that makes financial sense: by becoming a compute utility for anyone who can pay the bill.

The Compute Bottleneck Is Now the Whole Story

For years, the AI narrative was dominated by model releases, parameter counts, and benchmark scores. OpenAI's releases of GPT-3, GPT-4, and GPT-4o drove waves of excitement and investment. Anthropic's Claude models carved out a reputation for long-context reasoning and safety. Google's Gemini pushed multimodal integration. But all of that innovation ultimately runs on the same substrate: electricity, silicon, cooling, and networking. And that substrate is now the binding constraint.

Data center construction across the United States is running into serious local opposition. The Gallup organization has surveyed communities near proposed AI data centers and found that roughly 70 percent of respondents oppose new construction in their area, citing concerns about power consumption, water usage for cooling, property values, and the massive physical footprint of these facilities. Power grids in regions like Northern Virginia, the Pacific Northwest, and Texas are straining under the load. Semiconductor supply chains remain tight even as NVIDIA ramps Blackwell production.

When Anthropic pays $15 billion a year for access to someone else's supercomputer, that is a market signal: the companies that control compute will control AI, at least in the near term. This is analogous to the cloud computing era of the 2010s, when Amazon Web Services built the infrastructure that allowed every startup to deploy globally without owning a single server. Except this time, the stakes are higher, the capex is larger, and the demand is more urgent. Traditional cloud providers — AWS, Google Cloud, Microsoft Azure — are racing to expand their GPU capacity, but even their gargantuan buildouts are struggling to keep pace with AI-driven demand.

Anthropic's Rise From Startup to Compute Kingmaker

Anthropic was founded in 2021 by Dario and Daniela Amodei, former OpenAI executives who wanted to build AI systems with a stronger safety orientation. Five years later, the company is on the verge of turning a profit and has locked in a multi-billion-dollar compute partnership with one of the most storied engineering companies on the planet. That trajectory has been almost impossibly fast, even by Silicon Valley standards.

The company's Claude model has carved out a distinct position in the market. Where OpenAI has chased consumer adoption with ChatGPT and its widely adopted API platform, Microsoft has embedded Copilot into the Windows and Office ecosystems at the operating-system level, and Google has pushed Gemini across Search, Workspace, and Android, Anthropic has focused heavily on the enterprise. Claude is widely regarded as the strongest option for long-context coding tasks, detailed document analysis, and complex reasoning chains. That reputation has translated into contract wins that now rival the revenue of some Fortune 500 companies.

The financials tell the story. Anthropic's revenue quadrupled from roughly $2.6 billion in 2024 to a projected $10.9 billion in 2026. The company has expanded its headcount rapidly, opened new offices in London and Singapore, and built out a partnership ecosystem that includes Amazon Bedrock for model hosting, Palantir for government contracts, and now SpaceX for compute. Even with the $15 billion annual compute spend, the company is moving toward profitability because the margin on served API requests is so high once the fixed infrastructure cost is amortized over enough usage. The economics of AI are brutal at the start and extremely lucrative at scale, and Anthropic appears to have executed that transition faster than almost anyone expected.

There is also an irony worth noting. SpaceX merged with Elon Musk's xAI earlier this year, creating a conglomerate that includes both Grok and the Colossus data center. In a very real sense, Anthropic is now paying its competitor's parent company billions of dollars to power its own competing model. That is the kind of tangled competitive dynamic that only emerges in industries undergoing rapid, capital-intensive growth. It is reminiscent of early cloud computing, where companies competed fiercely on application logic while renting the same underlying servers from Amazon.

Google Pushes Gemini Into Your Living Room — And Your TV Remote

While the AI industry sorts out its data center drama, Google is busy deploying Gemini into consumer hardware. This week, the company announced a major upgrade to Gemini on Google TV, bringing two of its most interesting generative models — Nano Banana and Veo — directly to your television. The update is rolling out first on select 2025 and 2026 TCL Google TV models in the United States, with exclusivity for TCL for 60 days.

Nano Banana is Google's on-device image generation model, designed to run efficiently on consumer hardware without requiring cloud round-trips. Veo is the company's text-to-video model, capable of creating short video clips from natural language prompts. Combined with Gemini's existing multimodal capabilities, this means you can now sit on your couch, speak to your TV, and have it generate AI images or videos based on what you described. You can pull family photos from your Google Photos library, apply AI-generated styles or scenes, and produce stylized slideshows of vacations, birthdays, or holidays — all without touching a remote control.

The more immediately useful feature is voice-controlled settings. Instead of hunting through maze-like smart TV menus layered across dozens of settings submenus, you can simply tell Gemini, "the screen is too dim" or "I can't hear the dialogue," and the assistant will adjust brightness or volume accordingly. Google is even teasing the possibility of conversational TV watching: ask for more context about an actor in a scene, request a summary of last season's plot developments, or get real-time sports updates integrated into the broadcast. It is a small quality-of-life improvement, but it is exactly the kind of ambient assistance that AI should make frictionless.

This is also a meaningful shift in Google's AI strategy across devices. After years of pushing Gemini primarily through Search, Workspace, and the Pixel phone, the company is now embedding the model into televisions, smart home displays, and other ambient screens. The goal is to make Gemini the default interface for interacting with screens — a move that would lock in Google's AI ecosystem across an enormous surface area of consumer hardware. If every Google TV in the world becomes a Gemini terminal, Google's data advantage in training its models grows correspondingly. It is a flywheel strategy, and the company has been building it for years.

Tesla's Robotaxi Is Real, But It Is Also Very Small

Elon Musk has promised for years that Tesla's Full Self-Driving system would usher in an era of autonomous robotaxis, available to half the US population by the end of 2025. He projected millions of autonomous vehicles reshaping urban transportation within a few short years, eliminating the need for personal car ownership in cities and generating billions in recurring revenue from a fleet of self-driving taxis. It is now June 2026, and The Verge reports that Tesla's autonomous ridehail service is dramatically smaller than predicted: just 59 vehicles in a handful of Texas cities.

The gap between Musk's public timeline and Tesla's actual deployment is one of the most persistent gaps in modern tech. Musk tends to optimize for narrative momentum, painting pictures of decentralized transportation networks operating at planetary scale. Reality, as usual, is more complicated. Autonomous driving at scale requires not just trained models but robust sensor suites, extensive regulatory approvals, liability frameworks, fleet management infrastructure, and public trust. Fifty-nine cars in Texas is a real operational service with paying customers, but it is a proof of concept, not a revolution — and the distinction matters enormously when investors and regulators are evaluating Tesla's claims.

The contrast with Waymo is instructive. Waymo's fully autonomous ridehail service has been operating in Phoenix, San Francisco, Los Angeles, and Austin for years, with thousands of vehicles deployed and millions of rides completed. Unlike Tesla's camera-first approach — the company argues that human-level driving can be achieved with cameras alone, processing visual data through neural networks — Waymo uses lidar, radar, and high-definition maps. It is a more conservative strategy, carrying higher upfront hardware costs but offering more reliable perception in edge cases: fog, rain, construction zones, unusual road geometry. And it is producing real commercial results. Waymo Premier offers priority pickups and ten percent cash back on every trip, signaling that the company believes its service is mature enough to compete with traditional ridehailing on price and convenience.

Tesla's approach, by contrast, leans heavily on over-the-air software updates and camera-based perception with an optional hardware add-on. The argument from Tesla's side is that its fleet-learning advantage — millions of cars on the road collecting training data — will eventually allow it to leapfrog competitors by training on vastly more real-world scenarios. But the timeline for that leap keeps slipping, and the gap between Musk's promises and Tesla's actual deployable fleet is becoming harder to defend against regulators and shareholders alike. The Texas pilot, however modest in size, is an honest operational milestone. It would serve the narrative better if the milestones leading up to it had been equally accurate.

The Wearable Biotech Wave: From UV Sensors to Heart Rate Journalism

Shift from AI and cars to the body itself, and a similar pattern emerges: consumer hardware is getting smarter, sensors are getting cheaper, and the data is getting more personal. Two stories this week illustrate where wearable biotech is heading — and neither involves AI labs or electric vehicles.

The first is the launch of the Gem smart jewelry pendant from The90. Priced at $299 — discounted to $199 at launch — the Gem is a wearable that continuously measures UVA and UVB exposure throughout the day. Based on your skin profile — Fitzpatrick skin type, recent tanning history, estimated sunscreen usage — the companion app tells you how much sun is safe to get and when you should reapply. It is a more stylish take on the concept L'Oréal introduced with its My Skin Track UV sensor years ago, but the market for personalized health wearables has grown considerably since then, fueled by rising awareness of skin cancer and the broader consumerization of preventive health.

Skin cancer is the most common cancer in the United States, with nearly five million cases treated annually, and UV exposure is the single most preventable cause. A device that makes invisible radiation visible and actionable — "you've hit your safe limit for today" — is a genuine public health tool wrapped in consumer jewelry. The challenge, of course, is accuracy and regulatory compliance. The90 will need to validate its sensor precision carefully, and the FDA has been paying closer attention to consumer devices that make health claims. But the concept is sound, and the timing is right.

The second story is lighter but arguably more culturally revealing. Oura Ring data showed that New York Knicks fans' baseline heart rates increased by an average of 3.7 beats per minute during Game 4 of the NBA playoffs, with elevated rates persisting throughout the entire match. The effect was so pronounced that many fans reportedly struggled to sleep afterward. It is a reminder that wearable sensors are now generating datasets at population scale — datasets that journalists and researchers can use to tell hyper-specific stories about human behavior, emotion, and physiology that would have required small, expensive clinical studies a decade ago.

The broader trend is unmistakable: the line between consumer electronics and medical devices is blurring. Continuous glucose monitors from companies like Dexcom have become mainstream accessories for diabetics and biohackers alike. Apple Watch ECG functions are used to detect atrial fibrillation. WHOOP and Oura track recovery, sleep staging, and now cardiac response to emotional events. UV sensors, hormone-tracking rings, blood pressure cuffs that sync to your phone — they are all becoming more capable, cheaper, and smaller. The FDA has been slowly approving more consumer-grade health devices, creating a regulatory pathway that did not exist five years ago. The demand is undeniable, the technology is mature enough to ship, and the incumbents in traditional healthcare are beginning to feel the disruption.

Meta's AI Glasses Reach Blind Veterans At Scale

Rounding out the hardware side of this week's tech news, Meta announced that it is donating its AI-powered smart glasses to more than 130,000 blind veterans in the United States. The glasses use built-in cameras paired with Meta's multimodal AI to describe the wearer's surroundings in real time — identifying objects, reading text on signs and menus, recognizing faces, and providing spatial awareness cues through spatial audio. For visually impaired users, this is genuinely transformative accessibility technology.

Meta has been iterating on its smart glasses partnership with EssilorLuxottica — the parent company behind Ray-Ban, Oakley, and other iconic eyewear brands — for several product generations. The current Ray-Ban Meta Smart Glasses are the first version that actually looks and feels like regular glasses, with prescription lens support and a weight and balance that users can wear for a full day. The AI layer has improved dramatically with each firmware update, with real-time scene understanding and conversational interaction now possible at near-instant speeds. Deploying the technology to 130,000 veterans is not just a goodwill gesture — it is a large-scale field test of how AI-assisted vision works for users who rely on it full-time, every day, in complex real-world environments.

The strategic angle for Meta is also worth noting. The company has bet heavily on smart glasses as its post-smartphone platform. By anchoring its smart glasses narrative in a high-empathy, high-impact use case like veteran accessibility, Meta builds social license for a product category that many consumers still find slightly uncanny. If AI glasses can normalize as assistive technology before they normalize as consumer social-media accessories, Meta's path to mass adoption becomes significantly smoother. The company has been clear that it believes the future of computing is wearable and ambient, and this veteran initiative is a concrete, measurable step toward that future.

The Threads Connecting AI, Cars, and Biotech

None of these developments exist in isolation. The same semiconductor supply chains that train frontier AI models are also powering Tesla's full-self-driving chips and the custom processors inside Oura's ring and The90's pendant. The same capital markets that are funding SpaceX's $12.7 billion AI buildout are also bankrolling electric vehicle startups and biotech sensor companies. The same consumer expectation for ambient, always-on intelligence — "the screen is too dim" — is driving both Gemini on Google TV and AI glasses that narrate the physical world for users who cannot see it.

This convergence matters more than any single product launch or quarterly earnings report. We are entering an era where AI is not an app you open — it is the substrate beneath everything: the car you ride in, the glasses you wear, the TV you watch, the ring on your finger, the data center humming in Memphis. The winners in this era will not necessarily be the companies with the best models or the flashiest hardware. They will be the ones who can integrate intelligence seamlessly across enough surfaces of daily life that users stop noticing it is there — which, of course, is exactly how the best technology should feel.

From SpaceX's Colossus to Meta's smart glasses, from Google TV to Tesla's Austin robotaxis, from UV-sensing pendants to heart-rate journalism, the through-line is infrastructure becoming invisible. This week's news is not about individual gadgets or model releases. It is about the long, expensive, unglamorous process of turning science fiction into boring background utilities. And that process is accelerating faster than most timelines predicted. The companies that understand this — that treat compute, hardware, and integration as the real products, not the AI demos — will be the ones that survive the shakeout and define the next decade of technology.