Fusion Power Arrives

The infrastructure layer is closing. After decades of building foundational technologies and legal frameworks, we're watching the shift from open platforms to competitive ecosystems built on top of them.

Realta Fusion's electricity generation from a fusion reaction represents this transition perfectly. Not commercial viability yet, but proof that fusion can move from physics experiment to power generation. The focus now shifts to who builds the winning reactor design, who controls fuel supplies, and who owns the regulatory pathway to deployment.

The same pattern appears in space, where Rocket Lab's $8 billion acquisition of Iridium signals the industry moving beyond launch services toward integrated satellite operators who control both rockets and orbital infrastructure. Launch capability is table stakes. The value is in the network layer above it.

Even internet governance reflects this shift. Vinton Cerf's retirement from Google next week closes an era when protocol designers could shape the internet's direction. That power now belongs to platform operators and regulators negotiating over the mature infrastructure Cerf helped create.

Meanwhile, Japan's $6.16 billion commitment to building domestic AI models shows nations treating foundation models as critical infrastructure, not open research. The era of shared foundations is ending. The competition for advantage is beginning.

Deep Dive

The Protocol Wars Are Coming Back for AI Agents

The push toward AI agents creates a window for new infrastructure standards, and Vinton Cerf's prediction that formal protocols will be necessary for agent interoperability suggests the next battleground in AI will look less like the LLM race and more like the early internet. The companies that define how agents communicate and transact could capture outsized value without building the most capable models.

This matters because the current LLM landscape concentrates power in model builders. But if agents from different providers need to work together, as Cerf argues, someone has to define the standards for that coordination. Natural language is too ambiguous for reliable agent-to-agent communication. Formal protocols solve this, but whoever controls those protocols gains leverage over the entire ecosystem built on top.

The timing creates an opening for startups. Incumbent AI labs are focused on model performance and racing to AGI. Few are thinking seriously about the boring infrastructure of agent coordination. Yet these standards will determine whether we get an open ecosystem where any developer can build agents that interoperate, or a fragmented landscape where agents only work within walled gardens. The protocol question also surfaces trust and verification challenges. If an AI agent commits your company to a transaction, what's the technical framework for ensuring both sides executed correctly? How do you handle disputes? These are solvable problems, but they require standardization.

For founders, this suggests opportunity in the infrastructure layer beneath agents. Not the agents themselves, but the pipes connecting them. For VCs, it's a reminder that infrastructure bets often have longer time horizons but can generate platform-level returns. The companies defining agent protocols today could become the Cisco and Verisign of the agentic economy.

The risk is moving too early. Agent adoption is still nascent, and premature standardization can calcify around the wrong abstractions. But the pattern is recognizable: when new computing paradigms emerge, protocol control matters more than most people initially realize.

Vertical Integration Wins in Space

Rocket Lab's $8 billion acquisition of Iridium signals that launch capability alone no longer provides defensible moats in space. The value has moved to integrated operators who control both access to orbit and the satellites generating revenue once there. This changes the calculus for space startups and their investors.

The deal gives Rocket Lab immediate revenue, a customer base, and spectrum rights that typically take years to secure. More importantly, it validates the thesis that space companies need to own the full stack. SpaceX understood this early, building Starlink to create demand for its own launches. Now Rocket Lab is following the same playbook, buying its way into satellite operations rather than building from scratch.

For space founders, this means pure-play launch companies face a ceiling. Investors want to see a path beyond selling rides to orbit. The winning pitch now requires a clear answer to "what happens after launch?" That could mean proprietary satellites, subscription services, or data products. But it can't just be better rockets at lower prices.

The capital requirements also increase dramatically. Rocket Lab is spending $8 billion to acquire existing satellite infrastructure rather than building it organically. Smaller companies without access to that capital face a choice: find a niche where vertical integration doesn't matter, or accept being acquired by someone building the full stack. The middle ground of focusing solely on launch shrinks.

For investors, the deal validates billion-dollar outcomes in space beyond SpaceX. But it also suggests that meaningful value creation requires operating at scale across multiple layers of the stack. Series A checks into single-purpose space hardware look riskier when the long-term winners need to own satellites, spectrum, ground stations, and customer relationships in addition to launch.

The broader implication is that space is maturing from infrastructure buildout into industry consolidation. The question shifts from "can you get to orbit?" to "what do you control once you're there?"

Location Data Just Got More Expensive

The Supreme Court's ruling that geofence warrants require probable cause raises the cost of building location-based services. The decision affirms that users retain Fourth Amendment protections over location data even when shared with third parties, which means companies collecting that data now face higher legal and operational risk.

For consumer apps, the ruling creates incentive to collect less location data or to architect systems that can't easily respond to law enforcement requests. That's good for privacy but complicates product features that depend on precise location tracking. Navigation, local discovery, and real-time logistics all get harder when you can't log continuous location histories. Some developers will handle this by moving to on-device processing that never sends location data to servers. Others will aggregate and anonymize more aggressively. Both approaches reduce the resolution available for personalization and analytics.

The enforcement risk also increases. Tech companies now face potential Fourth Amendment challenges when they hand over location data without ensuring proper warrants were issued. That creates liability and forces legal review of law enforcement requests that previously got rubber-stamped. Smaller companies without robust legal teams are at a disadvantage.

For startups building location-dependent products, this means factoring legal cost and compliance complexity into early product decisions. Features that seemed straightforward now require privacy engineering resources. That's especially challenging in logistics, delivery, and real-time tracking applications where location precision drives core value.

The ruling also advantages end-to-end encrypted location solutions where the service provider genuinely cannot access user data. Signal and iMessage already work this way for messages. Expect similar architectures for location services, where user devices handle computation locally and servers only see encrypted blobs. That's technically harder but legally cleaner.

The broader trend is toward treating location data as sensitive by default. Companies that built products assuming freely available location histories will need to adapt. Those designing new products have the advantage of building privacy protections from the start.

Signal Shots

Payment Networks Launch Stablecoin to Break Circle's Model : Visa, Mastercard, and 140 other firms launched Open USD, a stablecoin that routes reserve interest to partners instead of a central issuer. The move directly attacks Circle's business model, which generated most of its revenue from interest on backing assets. This matters because it transforms stablecoins from a duopoly into open infrastructure competition. Watch whether Circle responds with pricing changes and how quickly OUSD gains adoption in corporate treasury operations. The launch also tests whether collective governance can move fast enough to compete with single-issuer platforms.

Nano-Infused Copper Cuts Grid Losses in Half : Arcturus emerged from stealth with technology that uses lasers to infuse carbon nanomaterials into copper, reducing electrical resistance and heat loss. The startup raised $8 million to scale from centimeters to tens of meters of wire for testing in motors and power distribution. Cutting grid losses in half would immediately unlock 3 percent more electricity capacity, equivalent to a year of demand growth. Watch for pilot deployments in data centers where efficiency gains directly reduce cooling costs. The challenge is scaling production while maintaining material properties at volume.

Amazon Builds Billion-Dollar Agent Deployment Team : AWS launched a forward-deployed engineering organization that will embed engineers inside companies to build custom AI agents, committing $1 billion in internal resources. The move follows similar initiatives from OpenAI and Anthropic, signaling that AI deployment has become a professional services business. This matters because it shifts AI value from models to implementation expertise. Watch how this affects cloud margins as AWS turns compute sales into labor-intensive consulting. The model also creates lock-in through custom integrations that become difficult to replace.

Etched Hits $5 Billion Valuation on Inference Chip Orders : AI chip startup Etched raised $500 million at a $5 billion valuation and reported $1 billion in contracted orders for inference systems built around its specialized chip. TSMC successfully manufactured the chip earlier this year and the company is now testing with customers. Inference cost and speed remain the biggest bottleneck for scaling AI applications, making specialized hardware attractive despite Nvidia's lead. Watch whether these systems deliver the claimed performance advantages in production environments. The surge in inference-focused chip startups suggests investors believe the GPU monoculture is breakable.

Tesla Tests Production Cybercabs Without Driver Controls : Tesla began testing production Cybercabs without steering wheels or pedals in Austin, following years of prototype testing with traditional controls. The tests use safety monitors in the passenger seat while the company waits for federal rules eliminating brake pedal requirements. This matters because it marks Tesla's first attempt at purpose-built autonomous vehicles rather than retrofitted consumer cars. Watch for crash data and edge case failures as the spotlight intensifies on vehicles that cannot be manually controlled. The gold two-seater design makes Tesla's struggles and successes far more visible than its existing Model Y robotaxi fleet.

ByteDance Commits $39 Billion to Brazil Data Center Complex : ByteDance is building a 1-gigawatt data center facility in a free-trade zone in Brazil's Ceará state, set to become its largest installation outside China. The $39 billion investment reflects the geographic distribution of AI infrastructure as compute demand outstrips domestic capacity. This matters because it shows Chinese tech companies building sovereign-scale infrastructure in markets where regulatory uncertainty is lower than the US or EU. Watch whether other hyperscalers follow into Latin America and how Brazil negotiates data residency requirements in exchange for the investment. The scale suggests ByteDance expects sustained compute needs beyond current model training cycles.

Scanning the Wire

Trump Drops Restrictions on Anthropic's Mythos and Fable Models : The administration's reversal on AI model controls leaves companies without clear guidance on what governs future releases. (TechCrunch)

Anthropic Launches Claude Sonnet 5 as Cheaper Agent Alternative : The new model brings stronger agentic capabilities and lower pricing, positioning itself as a more affordable option than Opus, GPT-5.5, and Gemini Pro. (TechCrunch)

Anthropic's Claude Science Builds Workbench for Computational Research : The tool consolidates databases, pipelines, and analysis into one environment, eliminating the need for scientists to switch between multiple platforms. (TechCrunch)

Wayve Launches $85M Employee Tender at $8.5B Valuation : The UK autonomous vehicle startup joins a growing trend of AI companies using liquidity events to compete for talent without going public. (TechCrunch)

Dish Files Chapter 11 After AT&T Spectrum Sale Delays : The company will continue operating Dish TV and Sling TV while winding down wireless operations following unforeseen holdups in its $23 billion 5G spectrum sale. (The Verge)

Getty Abandons $3.7B Shutterstock Merger After UK Regulator Blocks Key Assets : The deal collapsed despite US antitrust clearance after the UK Competition and Markets Authority imposed restrictions that would exclude part of Shutterstock's business. (The Verge)

Google Kills Tenor GIF API, Forces Platform Migrations : Third-party services including X and Discord must find alternative GIF providers while Tenor continues serving Google's own applications. (Ars Technica)

Neon Acquires OpenAI Documentary After Amazon Walks Away : The distributor purchased "Artificial," focused on Sam Altman and OpenAI, after Amazon dropped the project following its investment in the company. (NYT Technology)

Chinese Carmakers Rush to Design Local AI Chips : BYD, Nio, and others are accelerating development of domestically produced semiconductors with AI functions as part of broader push for chip self-sufficiency. (Financial Times)

South Korea Chip Exports Hit Record $44.82B in June : Total exports reached an all-time monthly high of $102.25 billion, up 70.9 percent year over year, driven by semiconductor demand. (Nikkei Asia)

Meta Launches $20 Monthly Premium Tier for Smart Glasses : Meta One Premium unlocks unlimited use of Conversation Focus and other AI features, while free users are limited to three hours per month. (The Verge)

MGX Closes $49B AI Fund Backing OpenAI and Anthropic : The Abu Dhabi-based investor's fund ranks among the largest ever raised for artificial intelligence, with stakes in OpenAI, Anthropic, and xAI before its SpaceX merger. (CNBC Tech)

Australia Sues Amazon Over Alleged Unfair Prime Terms : The competition regulator claims Amazon required Prime subscribers to pay an additional AU$2.99 monthly fee to avoid advertising without offering refunds for the change. (CNBC Tech)

Outlier

Australia Tests Contract Law Against Subscription Dark Patterns : Australia's competition regulator is suing Amazon for allegedly forcing Prime subscribers to pay AU$2.99 more to avoid ads without offering refunds. The case treats subscription changes as potential contract violations rather than just consumer protection issues. This signals a regulatory strategy that could make post-purchase modifications to digital services legally riskier than physical goods. If the approach spreads, expect subscription businesses to front-load terms more explicitly or face liability for mid-contract feature changes. The shift would force companies to treat digital subscriptions more like traditional service contracts with change-of-terms protections rather than perpetually mutable agreements.

The internet's founding architect retires the same week fusion produces electricity and rockets merge with satellites. Turns out the infrastructure era doesn't end so much as calcify into something worth fighting over. See you next week when we'll find out who's winning.