The Defense Tech Awakening

The most expensive problem in technology right now is not building AI systems. It is reorganizing companies that were not built for them. Meta's forced migration of 1,000+ engineers into AI units ahead of 10% layoffs reveals something critical about the current transition: legacy organizational structures designed for social graphs and engagement optimization do not naturally accommodate the demands of frontier AI development.

Meanwhile, companies architected from first principles for today's technological reality are capturing unprecedented capital and market position. Anduril's $5 billion raise at a $61 billion valuation and Cerebras's $5.55 billion IPO represent more than defense and infrastructure success stories. They signal investor conviction that greenfield companies, unburdened by legacy systems and cultures, can execute faster than incumbents can transform.

The enterprise data supports this thesis. Anthropic now serves more business customers than OpenAI, according to expense data from Ramp, a remarkable achievement for a younger lab. This is not just about product quality. It reflects something deeper: organizational buyers are hedging against the transition costs they see in their own operations, diversifying away from the perceived complexity of OpenAI's consumer-to-enterprise evolution.

The tax on legacy is becoming measurable, and it is being paid in both capital efficiency and market position.

Deep Dive

Enterprise Buyers Are Rewriting the AI Competitive Map

The enterprise AI market is producing different winners than the consumer market, and the gap is widening. Anthropic now serves more business customers than OpenAI according to expense data from fintech firm Ramp, marking the first time the younger lab has held the top position. Among Ramp's 50,000+ client companies, 34.4% pay for Anthropic services versus 32.3% for OpenAI.

This reversal has been building for months. In May 2025, only 9% of businesses paid for Anthropic products. That share climbed 26 percentage points over 12 months while OpenAI's declined 1%. The pattern extends beyond Ramp's dataset. On OpenRouter's leaderboard, which samples a different user base, OpenAI last ranked above Anthropic in December 2025.

What is driving this shift matters more than the shift itself. Anthropic executed a technical-first strategy: start with demanding enterprise customers, solve their specific needs around safety and reliability, then expand outward. This approach compounds. Technical buyers become internal champions. Their teams build dependencies. Switching costs accumulate. OpenAI's trajectory moved in the opposite direction, building massive consumer adoption through ChatGPT then working backward to enterprise needs.

The implications for founders are direct. Consumer virality and enterprise durability require different product architectures, go-to-market motions, and organizational capabilities. Companies that try to bolt enterprise features onto consumer products face structural disadvantages against competitors built for business buyers from day one. For VCs, this dynamic suggests that large consumer user bases may be weaker moats in AI than previously assumed. Enterprise relationships, particularly in regulated industries and technical workflows, appear to be creating more defensible positions. The market is rewarding focused execution over platform ambitions, at least for now.

Defense Tech Capital Formation Has Reached Industrial Scale

Venture capital has moved beyond dabbling in defense. Anduril's $5 billion raise at a $61 billion valuation, more than double its $30.5 billion valuation from less than a year ago, represents the emergence of a fully formed capital cycle in defense technology. This is not experimental allocation. This is core portfolio strategy for top-tier firms like Andreessen Horowitz and Thrive Capital.

The financial model is proving itself. Anduril doubled revenue to $2.2 billion in 2025, demonstrating that defense contracts can scale predictably once a company breaks through initial procurement barriers. The company has now raised over $11 billion total, putting it in the same capital formation category as foundational AI labs. Similar patterns are emerging across the sector: Shield AI raised $1.5 billion at a $12.7 billion valuation in March, while European defense tech company Helsing is reportedly close to raising $1.2 billion at an $18 billion valuation.

What changed is structural, not cyclical. The Department of Defense is actively fostering competition among new contractors. When the Air Force recently selected Shield AI's software to work with Anduril's Fury autonomous fighter jet, it signaled a clear strategy: avoid single-vendor lock-in even with rising-star startups. This creates better venture outcomes by enabling multiple winners rather than forcing a winner-take-all dynamic.

For founders, the lesson is timing. Defense procurement is notoriously slow, but once vendors prove themselves, contract expansion can be dramatic and sustained. The capital intensity required to reach that proof point has dropped as manufacturing and autonomy technologies mature. For VCs, defense tech now offers the same risk-return profile that made enterprise SaaS attractive: long sales cycles offset by high retention, predictable growth, and enormous TAM once contracts scale. The sector has graduated from contrarian bet to institutional asset class.

The Hidden Tax of Legacy Shows Up in Employee Tracking Software

Meta's decision to install monitoring software on US employee computers to train AI models has crystallized a larger problem facing incumbents: they must now tax their existing operations to fund their transformation. The company is cutting 10% of its workforce, roughly 8,000 people, while simultaneously forcing over 1,000 top engineers into a new Applied AI Engineering division. Those who refused the transfer faced layoffs, an unusual threat that several employees described as a "draft."

This is not just a morale problem, though morale is reportedly at historic lows. This is a capital allocation problem made visible through organizational stress. Meta is spending between $125 billion and $145 billion on capital expenditures this year, largely for AI infrastructure, while cutting compensation for the second consecutive year. Median total compensation fell from $417,400 in 2024 to $388,200 in 2025. The contrast between cost-cutting for existing employees and reported $100 million annual offers for top AI researchers illustrates the reallocation underway.

The monitoring software, called Model Capability Initiative, tracks employee typing and clicking to generate training data. It cannot be opted out of, according to current employees. That it was deployed only in the US, not Europe, suggests the company recognizes the practice would not survive scrutiny under stricter labor protections. Employee groups have begun organizing protests, with some UK workers pursuing formal union representation.

For tech workers, this creates a clear calculus: optionality matters more than brand. Many employees reportedly hope to be laid off to receive 16 weeks severance and 18 months of paid healthcare rather than remain in what they describe as an eroding culture. For founders, Meta's struggles offer a preview of transformation costs at scale. Organizations built for one technological paradigm do not easily reshape themselves for another. The greenfield advantage is real, and it is accelerating.

Signal Shots

Anthropic Targets 36 Million Small Businesses: Anthropic launched Claude for Small Business, a suite integrating with QuickBooks, Canva, and PayPal to automate bookkeeping and marketing for local companies. The AI platform wars are expanding downmarket: while enterprise adoption drove early growth, the real volume play is the 36 million small businesses representing 44% of US GDP. Watch whether SMBs adopt AI faster than they did cloud software, and whether simplified interfaces can overcome the training barriers that have kept smaller firms behind. OpenAI launched similar offerings in 2023, but Anthropic is betting on a 10-city roadshow to build local adoption.

Geothermal IPO Pops on Data Center Demand: Fervo Energy raised $1.89 billion in an upsized IPO and saw shares jump 33% to a $10 billion valuation, driven by demand for AI data center power. Enhanced geothermal can provide 24/7 baseload power that data center operators will pay premium rates to secure. The company has cut drilling costs and time by two-thirds across 14 wells, with its Cape Station project permitted for up to 4 gigawatts. Watch whether Fervo can scale interconnection capacity or pivots to behind-the-meter deals with hyperscalers seeking dedicated power sources.

Cisco Surges on $9 Billion AI Infrastructure Pipeline: Cisco shares jumped 17% after reporting $5.3 billion in AI infrastructure orders year-to-date and raising full-year expectations to $9 billion, while cutting 4,000 jobs. The networking incumbent is finally capturing AI buildout spending after trailing cloud-native competitors. CEO Chuck Robbins framed the cuts as reallocation toward high-demand areas, a pattern emerging across tech: companies are taxing legacy operations to fund AI transformation. Watch whether Cisco's new silicon and routing products can sustain market share gains, and whether competitors face similar workforce restructuring to redirect investment.

Tech Hiring Tilts Senior as AI Reshapes Roles: IT and CS job postings rose 14.2% year-over-year in April, but entry-level roles fell from 8.1% to 7.4% of openings while senior positions grew from 38.8% to 43.1%, according to ZipRecruiter. Companies are seeking experienced engineers who can deploy AI agents and integrate new tools, not train junior developers on fundamentals being automated away. This inverts the traditional hiring pyramid and creates a two-tier market. Watch whether bootcamps and universities adapt curricula faster than the skill gap widens, and whether companies invest in upskilling mid-career engineers or simply bid up senior talent.

US Approves Chinese Access to H200 Chips, But No Deliveries: The US cleared roughly 10 Chinese companies, including Alibaba, Tencent, ByteDance, and JD.com, to purchase Nvidia H200 chips, yet no deliveries have occurred. The approval signals willingness to allow some AI development in China while maintaining practical control through implementation delays. This creates uncertainty for Chinese labs planning infrastructure buildout and advantages US-based competitors. Watch whether deliveries begin or approvals remain symbolic, and whether Chinese firms accelerate domestic chip alternatives or shift more compute to overseas facilities.

Google Preps Gemini Release Below Frontier Level: Google will announce a new Gemini model at its I/O conference next week that sources say lands roughly at GPT-5.5 level, well short of Anthropic's Mythos. This marks Google's continued position as a fast follower rather than frontier leader, despite massive compute advantages. The release timing suggests Google is prioritizing shipping cadence over benchmark leadership, possibly to maintain developer and enterprise engagement. Watch whether Google can articulate a differentiated value proposition beyond cost and integration, or whether its AI strategy becomes primarily about distribution through Search and Workspace rather than model superiority.

Scanning the Wire

Rivian Spinoff Mind Robotics Raises $400M: The autonomous vehicle company, first revealed in late 2025, has now raised more than $1 billion total as it scales production of robotics systems originally developed for Rivian's manufacturing operations. (TechCrunch)

Fractile Raises $220M for AI Inference Chips: The London startup designing chips that place compute and memory on the same die attracted investment from Factorial Funds, Accel, and Founders Fund as it moves toward production with Anthropic reportedly in early customer discussions. (WSJ)

Palo Alto Networks Warns AI Cyberattacks Becoming Normal: The company found and fixed 75 flaws this month versus its usual five as increasingly sophisticated AI models enable faster and more automated attacks that pressure security teams to accelerate their defense strategies. (CNBC)

Lawmakers Demand Answers on Canvas Data Breaches: House members want to know how hackers broke into education tech giant Instructure twice and stole student data from Canvas, the flagship software used by universities and schools nationwide. (TechCrunch)

Jensen Huang Foundation Donates $108M in AI Computing: The Nvidia CEO's foundation bought computing time from CoreWeave and is distributing it to universities and nonprofit research institutes to expand academic access to frontier AI infrastructure. (Reuters)

FCC Approves AT&T and Starlink Purchase of EchoStar Spectrum: The commission angered small carriers by helping the telecom giants acquire wireless licenses, with the chair having pressured EchoStar to sell. (Ars Technica)

Origin Lab Raises $8M to Broker Video Game Data Sales: The startup will operate a marketplace where AI labs can license high-quality training data from video game companies building world models and simulation environments. (TechCrunch)

Windows Update to Auto-Rollback Faulty Drivers: Microsoft is adding a feature that automatically reverses problematic driver installations pushed through Windows Update, part of a broader effort to reduce update-related disruptions. (The Verge)

Instagram Tests Instants Feature: The new offering combines elements from Snapchat and BeReal, letting users share disappearing photos with close friends or mutual followers that can be viewed once and remain available for 24 hours. (TechCrunch)

Cisco Cuts 4,000 Jobs, Offers Free Cisco Training: The networking company is reducing headcount while providing affected employees with training on Cisco products, framing the cuts as part of a shift toward AI and cloud-native infrastructure priorities. (The Register)

Outlier

Cisco Trains the Laid Off to Sell Against Cisco: Cisco is cutting 4,000 jobs while offering those same workers free training on Cisco products, a bizarre arrangement that amounts to funding your future competitors' expertise. The company frames this as generosity, but the actual signal is darker: tech incumbents have become so desperate to appear humane during AI-driven restructuring that they will subsidize the exact knowledge transfer that undermines their market position. This hints at a future where large companies cannot hold talent during transitions and instead resort to controlled dispersal strategies, effectively open-sourcing their institutional knowledge to maintain employer brand. The severance package as competitive disadvantage.

The severance package as competitive disadvantage might be the most honest description of corporate strategy we'll see all year. If the future of work is paying people to leave with exactly the skills needed to compete against you, at least the job market will stay interesting.