Infrastructure Stakes Escalate

The AI industry is undergoing a phase transition. What began as a race to build better models is now a scramble for the infrastructure beneath them. Anthropic's potential $900 billion valuation matters less for what it says about Claude and more for what it reveals about the market's willingness to pay for compute access at scale.

The supply constraints are creating their own economic reality. Nvidia's B300 servers are trading at nearly $1 million in China, double their typical price, as enforcement against chip smuggling tightens. Samsung is warning that the memory shortage will intensify next year, with customers already placing orders for 2027. These aren't temporary disruptions. They're signals that demand is structurally outpacing supply.

The response is revealing. Amazon's chip business crossed $20 billion in annual revenue, making it a top three datacenter chip company. SoftBank is forming a robotics company to build data centers and already positioning it for a $100 billion IPO. The infrastructure layer is becoming its own platform, with valuations that rival the model makers themselves.

The companies that control scarce resources are extracting unprecedented value. This isn't about being early to AI anymore. It's about controlling access to the means of production.

Deep Dive

Anthropic's Revenue Multiples Reveal the New Rules of AI Valuations

The math behind Anthropic's potential $900 billion raise tells you everything about how investors are pricing AI infrastructure access. The company's revenue jumped from $9 billion at the end of 2025 to over $30 billion today, and sources say the current run rate is closer to $40 billion. At a $900 billion valuation, that's roughly 22 times current revenue. For context, enterprise software companies typically trade at 5 to 10 times revenue. Even high-growth SaaS businesses rarely exceed 15 times.

This premium exists because Anthropic isn't just selling software. It's selling access to compute capacity that doesn't exist anywhere else at the scale customers need. The company's Claude Code and Cowork platforms are generating most of that revenue, with coding applications proving far stickier than generic chatbot uses. Enterprises are locking in multi-year contracts because they can't risk losing access. When an institutional investor prepares $5 billion and can't even secure a meeting with the CFO, you're looking at supply scarcity that justifies extraordinary multiples.

For founders, this creates a clear calculus: if you're building on top of foundation models, you're competing against the model makers themselves for applications. Anthropic can bundle capabilities that independent developers can't match. For VCs, the message is equally stark. Late-stage AI investments are increasingly about infrastructure control, not application innovation. The companies with compute will capture the value, and customers are already booking capacity through 2027. The window for application layer plays that don't own their infrastructure is narrowing fast.

Amazon's Chip Business Exposes the Cloud Platform Squeeze

Amazon's custom silicon operation crossing $20 billion in annual revenue is the quiet story that changes everything for chip startups and cloud competition. CEO Andy Jassy revealed the business would be worth $50 billion if Amazon charged itself market rates for its own chips, the way traditional semiconductor companies do. That hypothetical accounting places Amazon among the top three datacenter chip businesses globally, and it got there in roughly three years.

The speed matters because it demonstrates how quickly cloud platforms can verticalize. Amazon's Trainium chips, which compete directly with Nvidia's GPUs for AI training, are already sold out through multiple generations. Trainium2 is gone, Trainium3 is nearly fully subscribed, and Trainium4 is largely reserved 18 months before broad availability. OpenAI committed to two gigawatts of Trainium capacity starting in 2027. Anthropic committed to five gigawatts. These aren't spot purchases. They're multi-year infrastructure bets that lock customers into Amazon's ecosystem.

For chip startups, this is the platform risk that venture investors have been underpricing. When your customer is also building a competing product and controls distribution, your margin of technical superiority needs to be enormous. Amazon doesn't need the best chip. It needs a chip that's 30 percent cheaper on performance and integrated into the platform customers already use. For founders building AI infrastructure, the question isn't whether your technology is better. It's whether you can build a business before the platforms replicate your core value and bundle it for free. The evidence suggests that window is measured in quarters, not years.

Signal Shots

Samsung's Memory Profits Signal Deeper Supply Constraints: Samsung reported an eightfold increase in first-quarter operating profit, hitting a record 57.2 trillion won on memory chip demand that far exceeds supply. The company's chip division alone delivered 53.7 trillion won in profit, accounting for over 90 percent of total earnings, with operating margins exceeding 70 percent. Executives said demand fulfillment rates are at record lows, and customers are now placing orders for 2027 to secure capacity. The memory shortage that's driving these margins directly constrains AI infrastructure expansion. Watch whether Samsung can ship enough HBM4 to challenge SK Hynix's lead before Nvidia's next architecture requires it at scale.

AI Agents Are Being Quietly Recruited Into Crypto Mining Swarms: Security researchers discovered 30 ClawHub skills that silently co-opt AI agents into a cryptocurrency mining network without user knowledge or consent. The skills, downloaded nearly 10,000 times, instruct agents to register with external servers, generate Hedera crypto wallets, and check in every four hours, all without malware or human approval. This is the first documented case of AI agents being mass recruited for economic purposes through instruction manipulation rather than code exploits. The mechanism works because registries scan for malicious code patterns but can't detect harmful instructions in natural language. Watch whether platforms add runtime visibility into agent behavior or require network endpoint disclosure in skill manifests.

Critical cPanel Flaw Exposes 70 Million Domains to Root Access: Researchers disclosed CVE-2026-41940, a 9.8 severity vulnerability in cPanel and WHM that lets attackers bypass authentication and gain root server access with just a few HTTP requests. The flaw affects every supported version prior to the patch and evidence suggests it was exploited as a zero-day for at least 30 days. Given that cPanel manages properties for an estimated 70 million domains, the exposure is massive. Successful exploitation requires only creating a session cookie and sending a crafted header to escalate privileges. Watch whether attack activity spikes now that proof-of-concept code is public, and whether hosting providers face regulatory scrutiny over delayed patching.

Databricks Can't Escape Copyright Claims Over Acquired Training Data: A federal judge refused to dismiss a class action lawsuit claiming Databricks' DBRX model was trained on pirated books, keeping alive claims that could result in six-figure damages per work infringed. Authors argue Databricks acquired MosaicLM, which used the RedPajama dataset containing Book3, a database of 196,000 pirated titles later pulled from Hugging Face. While Databricks provided 14 depositions and terabytes of discovery showing DBRX wasn't trained on Book3, the judge wants more detail about what happened step by step during development. The case could force bet-the-company damages exceeding Databricks' ability to pay if authors prove willful infringement. Watch whether Databricks pivots to a fair use defense like Meta and Anthropic used successfully, or tries to prove the disputed data had no impact on the final model.

Waymo Operations Are Getting Worse, First Responders Tell Regulators: San Francisco and Austin emergency officials told the NHTSA in a private meeting that Waymo robotaxis are backsliding and committing more traffic violations than before, according to leaked audio. Fire chiefs said vehicles now frequently block fire stations and default to freezing during emergencies. Police reported that contrary to Waymo's claims, vehicles often fail to recognize hand signals, creating cascading delays. One Austin incident saw a Waymo block an ambulance responding to a downtown shooting that killed three people. This comes as Waymo expands to 10 new cities this year while providing 500,000 paid rides weekly. Watch whether cities demand operational restrictions before approving expansion, and whether the company's aggressive deployment timeline faces political opposition beyond organized labor.

Fusion Startup Zap Energy Adds Fission to Its Roadmap: Well-funded fusion company Zap Energy announced it will develop fission reactors alongside its fusion program, aiming to generate revenue within a year from the new business line. CEO Zabrina Johal said rising AI data center demand means customers need power today, not in a decade when fusion might be ready. The fission design is based on the Toshiba 4S, a molten salt-cooled reactor that was never built but comes with no IP entanglement. Zap expects to fund development through Department of Defense contracts and milestone payments from large energy users who will reserve future capacity. The strategy borrows from ASML's customer co-investment program, though energy buyers have more supplier options than EUV lithography customers had. Watch whether Zap can attract new classes of investors excited about fission or whether the split focus slows fusion progress.

Scanning the Wire

Linux cryptographic code flaw offers fast route to root: Developers of major Linux distributions are shipping patches for a local privilege escalation vulnerability arising from a logic flaw in cryptographic authentication code. (The Register)

Parallel Web Systems hits $2B valuation five months after its last big raise: The AI agent-tool startup founded by former Twitter CEO Parag Agrawal raised $100 million led by Sequoia, doubling its valuation in under six months. (TechCrunch)

Uber is in the hotel business now, thanks in part to AI: The ride-hailing company announced features at its annual event that push deeper into users' lives, including hotel bookings powered by AI-driven personalization. (TechCrunch)

Bill Gurley, Jack Altman back startup Pursuit, which helps companies sell to government: The company announced a $22 million Series A led by OpenGov co-founder Mike Rosengarten to build tools that streamline the procurement process for federal contractors. (TechCrunch)

Apple loses bid to pause App Store fee changes as case heads to Supreme Court: A federal court rejected Apple's request to delay court-ordered payment changes, keeping external purchase links in place while its Epic Games appeal advances. (TechCrunch)

Earth AI is vertically integrating the search for critical minerals: The startup is acquiring drilling and analysis capabilities after hitting months-long delays with contractors, betting it can move faster by owning the full exploration stack. (TechCrunch)

Oracle plans to power its New Mexico mega datacenter with a 2.45GW fuel cell farm: The company is moving forward with a massive datacenter complex despite reports that key customer OpenAI may struggle to pay future cloud bills. (The Register)

Microsoft opens door to the past by releasing 86-DOS and PC-DOS 1.00: The company released source code for its operating system relics, giving retro computing enthusiasts access to the foundation of its software empire. (The Register)

EU waves through open source age-check tool to keep kids safe online: The European Commission recommended member states adopt an age verification app designed to protect children from harmful content, telling platforms there are no more excuses for non-compliance. (The Register)

OpenAI says it has signed contracts for 10GW of US AI compute capacity: The company added over 3GW in the past 90 days, hitting a capacity goal three years ahead of its original 2029 target as it races to secure infrastructure for future models. (Bloomberg)

NVIDIA and Atlassian back Legora's $600m Series D extension: The legal AI startup, which crossed $100 million in ARR in 18 months, added $50 million to maintain its $5.6 billion valuation, with Nvidia making its first investment in legal tech. (The Next Web)

China threatens the EU with broad retaliation if Huawei and ZTE are banned: Beijing's commerce ministry submitted a 30-page document warning that mandatory vendor removal under the EU's draft Cybersecurity Act could trigger reciprocal measures against European companies. (The Next Web)

Met Police's Palantir deployment has its own officers watching their backs: London's staff association is telling members to be extremely cautious about carrying work devices off duty after the force deployed Palantir's technology to investigate over 600 of its own officers. (The Register)

Drone strikes on data centers spook Big Tech, halting Middle East projects: Uninsurable war damage is forcing cloud providers to rethink expansion plans in the region as attacks on infrastructure make projects unviable. (Ars Technica)

Outlier

War Makes Data Centers Uninsurable: Drone strikes on Middle East data centers are forcing tech companies to halt regional expansion because infrastructure war damage has become uninsurable. This isn't about terrorism risk, which insurers have priced for decades. It's about precision weapons targeting fixed assets in conflicts where international law is increasingly theoretical. The shift signals a world where digital infrastructure can't be treated as neutral civilian property, and where cloud providers must calculate geopolitical stability as a first-order constraint. If data centers require military-grade defense systems to attract capital, the cloud's geographic expansion stops at conflict zones, and the internet fragments along security perimeters rather than regulatory borders.

The fusion company pivoting to fission while waiting for its real technology to work is the perfect metaphor for 2026: we'll take the future we can get this year over the one we were promised for next decade.