U.S. AI Hawks Face a Revolt From Silicon Valley

Gemini's user base nears a billion, Alphabet doubles down on $200 billion in AI capex, and the campaign to contain Chinese open-weight models loses support from America's own chipmaker and startup ecosystem.

July 23, 2026 | Reading time: 11 minutes | Issue #219

Lead

Google said on Tuesday that Gemini now has 950 million monthly users, up from 750 million in February. Alphabet, its parent, reported second-quarter revenue of $119.8 billion, a 24 percent increase year on year, and raised its 2026 capital expenditure guidance to $195 billion–$205 billion. The scale is the story: a consumer AI product is approaching one billion users while the company behind it commits roughly the annual GDP of Greece to building the infrastructure to serve it.

The earnings figure landed the same day Treasury Secretary Scott Bessent warned on Fox Business that the administration could sanction Chinese AI models it believes were distilled from American ones. "We are finding watermarks of our U.S. large language models on many of the Chinese models," Bessent said. The administration also plans to hold AI talks with China in September, with Bessent representing the U.S.

The containment push is meeting internal resistance. Nvidia CEO Jensen Huang told Axios the same day that American companies should "absolutely" be allowed to use Chinese open-weight models and that the campaign to ban them is the real threat. "There's no scenario where China runs U.S. companies off the road," he said. "Zero possibility." His argument is commercial: cheaper, open models expand adoption, and adoption expands demand for chips, data centers, and computing power — all products Nvidia sells.

The conflict is now structural, not ideological. Washington is treating open-weight diffusion as a national security problem. Nvidia is treating it as a market-expansion mechanism. Startups are treating it as a cost floor they cannot afford to lose. The result is a three-way split between policymakers, infrastructure vendors, and application builders over what "American AI leadership" actually means.

Briefs

Reddit and Google Renegotiate Their $60 Million AI Data Deal

Reddit shares fell 8 percent on Wednesday after The Information reported that the two companies have discussed Google losing access to Reddit content for AI training. Reddit and Google signed a $60 million-per-year deal in 2024. That agreement is now up for renewal against a backdrop of AI-generated search summaries that have reduced referral traffic to publishers and platforms. A Reddit spokesperson told CNBC the company is approaching negotiations "just like any business should, by focusing on doing what's best for Reddit." The case is becoming a template: content platforms that once sold AI licensing deals are now reassessing whether the traffic and brand trade-offs are still worth it.

Sources: CNBC

Travis Kalanick's Robotics Holding Company Raises $1.7 Billion

Travis Kalanick's Atoms raised $1.7 billion in a Series A led by Andreessen Horowitz, with Bain Capital, Fifth Wall, and Uber also joining. Atoms is a rebranded holding company built atop Cloud Kitchens and Pronto, Kalanick's autonomous-vehicle play for mining and industrial yards. Ben Horowitz is joining the board. Kalanick described the firm's goal as building a "wheelbase for robots." The round reconnects Kalanick with Uber, the company that removed him as CEO in 2017, and signals that a16z is willing to bet on physical-world automation at scale even as software-only AI startups face margin pressure.

Sources: TechCrunch

OpenAI and Anthropic Lobbying Spending Hit a Record in Q2

The two leading AI labs spent a combined $3.17 million on federal lobbying in the second quarter of 2026, up 23 percent from the first quarter and a record for both. Anthropic spent $1.97 million, more than Nvidia and nearly matching Oracle. OpenAI spent $1.2 million, an 18 percent quarterly increase. The disclosures show both companies lobbying on cybersecurity, copyright, cloud computing, and defense issues. While the totals are still small next to legacy tech and defense budgets, the growth rate signals how much policy risk the labs now see in Washington.

Sources: CNBC

Google Adds a Cybersecurity-Focused Gemini Model

Google expanded its Gemini lineup with Gemini 3.5 Flash Cyber, a version trained to find and fix software vulnerabilities. The model will initially be restricted to governments and trusted partners. Google also released Gemini 3.6 Flash, which uses up to 17 percent fewer tokens than the previous generation, and Gemini 3.5 Flash-Lite for faster, high-volume workloads. The launch came a day before Alphabet's earnings report and is best read as a defensive move: Anthropic's Claude Mythos has been marketed as a restricted cyber model, and Google is offering a competing capability while undercutting on price and access.

Sources: CNBC

Compute Watch

TSMC's U.S. Expansion Will Raise Chip Prices

TSMC reported second-quarter profit up 77.4 percent year on year, but the company warned that overseas expansion is diluting margins. CFO Wendell Huang said gross margins will be diluted by 2 to 3 percentage points in the early years of overseas fab ramp-up, widening to 3 to 4 points later. Morningstar estimates that chips produced in the U.S. will cost 20 to 50 percent more than equivalent wafers from Taiwan. Nikkei reported that TSMC intends to raise prices for advanced and mature nodes by as much as 10 percent starting in 2027. Because TSMC has no material competition at the leading edge, much of that increase is expected to pass to customers such as NVIDIA, AMD, Apple, and cloud providers. The geopolitical push to reshore semiconductor manufacturing is producing a price shock that the AI build-out will absorb.

Sources: CNBC, Nikkei Asia

Builder's Corner

Startups Race to Build the Payments and Kernel Layers for Agents

Two infrastructure startups raised capital this week for the plumbing that AI agents need but incumbents have not built. Natural, founded by former Ivella executives, raised $30 million in a Series A led by Forerunner to build an agent orchestration layer for payments. The company is positioning itself against Stripe, arguing that existing financial rails were designed for human-initiated transactions and cannot support autonomous vendor payments, collections, and reconciliation. Infinity raised $15 million at a $100 million valuation from Touring Capital, Principal VC, and researchers from OpenAI and Anthropic. Its agent, Ignition, writes the low-level kernels and inference code needed to run AI models on chips that compete with Nvidia. Customers include D-Matrix. The two companies operate at opposite ends of the agent stack — one at the money layer, one at the silicon layer — but both are betting that the established incumbents are too slow to adapt to autonomous software.

Sources: TechCrunch - Natural, TechCrunch - Infinity

Eastern Front

Beijing's Open-Weight Push Splits Washington and Silicon Valley

The Trump administration's effort to contain Chinese AI models is colliding with the reality that Western companies are already dependent on them. Treasury Secretary Bessent said the administration would investigate "watermarks" of U.S. models inside Chinese open-weight systems and could sanction them for model "theft." Hours later, Nvidia's Jensen Huang said American companies should be allowed to use those models and that restricting them would backfire.

The divide is not limited to Washington and chipmakers. A newly formed Little Tech Association representing roughly 200 startups, including Y Combinator and Proton, sent letters to President Trump and Commerce Secretary Howard Lutnick warning that cutting off access to Chinese open-weight models would "instantly" kill hundreds of U.S. companies. Mira Murati's Thinking Machines, which raised $2 billion, is reportedly using Alibaba's Qwen as part of its work. Apple has received Chinese regulatory approval to use Baidu's Ernie and other local models on iPhones in China. Rest of World described the dynamic bluntly: Chinese frontier models are becoming core infrastructure that global technology companies are building on, even as the U.S. tries to wall off its own frontier.

Sources: CNBC - Bessent, Axios - Huang, Politico - Little Tech, Rest of World

India Lens

Glow's AI-Native Endpoint Security Has a Large Israeli-Indian Engineering Base

Glow, the Palo Alto cybersecurity startup that emerged from stealth with a $1.2 billion valuation, employs nearly 100 people, about 70 percent of them in Israel and the rest in the U.S. The company's engineering model — a small U.S. headquarters with a deep technical center in Israel — reflects how AI-native security startups are distributing talent across global technical hubs rather than concentrating it in San Francisco. While Glow does not have an India base, the pattern is relevant to India's AI services sector: as endpoint security, chip-software automation, and agent orchestration become core AI infrastructure categories, the value is shifting from generic cloud migration work to specialized AI engineering. Indian IT services firms such as TCS have announced plans to train thousands of AI deployment engineers; the question is whether they can move fast enough to capture the kernel, security, and agent-infrastructure work that companies like Infinity and Glow are building from scratch.

Sources: TechCrunch - Glow

Europe

EU Orders Google to Open Android and Search to AI Rivals

The European Commission told Google this week that it must open Android and Search to rival AI services, a decision that builds on the Digital Markets Act. The Commission wants Google to share search data and make Android more interoperable with competing AI assistants. Google said it would appeal. The ruling comes as U.S. AI labs are lobbying Washington to restrict Chinese models while European regulators are trying to restrict U.S. model dominance on the continent. The contradiction is becoming harder to ignore: the same American companies arguing for open global access to AI are fighting forced openness in Europe.

Sources: The Verge

The View

Today's stories trace three overlapping fights over who controls the AI stack. The first is geopolitical: the Trump administration is trying to contain Chinese open-weight models while U.S. companies keep building on them. That gap between policy and practice is not a communications problem; it is a recognition that Chinese models are now price-competitive infrastructure. The second fight is economic: Alphabet's $195 billion–$205 billion capex budget and TSMC's planned 10 percent price hike mean the cost of compute is rising even as open-weight models drive API prices down. The squeeze lands on application-layer margins. The third fight is structural: Nvidia, startups, and the AI labs want different things from Washington. Nvidia wants open diffusion because it expands hardware demand. Startups want cheap open models because they cannot survive on frontier API pricing. The labs want regulatory protection from distillation while they build closed frontier systems.

These fights are linked. The more successful Chinese open-weight models are at compressing the price of intelligence, the harder it becomes for U.S. labs to maintain premium API margins and the more urgent their lobbying becomes. At the same time, cheaper models increase demand for the chips and data centers that Alphabet and Nvidia are racing to supply. The result is a strange equilibrium where openness and concentration are advancing together.

The Miss

A paper on arXiv this week argues that small, open-weight language models can be orchestrated to outperform a single large LLM on malware analysis. The authors combine models with different specializations and a routing layer, achieving better results at lower cost on malware classification and explanation tasks. The finding matters because it undercuts the assumption that security-critical tasks require closed frontier APIs. If open-weight orchestration works for malware, it works for many other enterprise workloads. Coverage: none outside the preprint server.

Sources: arXiv

Pull Quotes

"These Chinese models are excellent. Open-source models that are excellent should be used." — Jensen Huang, CEO, Nvidia, Axios

"There's no scenario where China runs U.S. companies off the road. Zero possibility." — Jensen Huang, CEO, Nvidia, Axios

"American leadership requires two things: world-leading American open-weight models and continued access for U.S. builders to open models already available worldwide." — Little Tech Association letter, Politico

"We are finding watermarks of our U.S. large language models on many of the Chinese models, and that's unacceptable." — Scott Bessent, U.S. Treasury Secretary, Fox Business

Reads & Links

• Alphabet Q2 2026 earnings release and Gemini user stat: Alphabet investor relations / The Verge
• Nvidia's Jensen Huang on Chinese open-source AI: Axios
• Treasury Secretary Bessent's sanctions warning on Chinese AI "theft": CNBC
• Little Tech Association letter opposing restrictions on Chinese open-weight models: Politico
• Rest of World on how Silicon Valley is already building on Chinese models: Rest of World
• TSMC price hikes and U.S. margin squeeze: CNBC / Nikkei Asia
• OpenAI and Anthropic lobbying disclosures: CNBC
• Small open-weight model orchestration for malware analysis: arXiv 2607.20216

The harder Washington pushes to contain open-weight diffusion, the more the builders it claims to protect build around it.