AI Market Hit by Sanctions Threats, Safety Breaches and $205 Billion Spending

Alphabet Inc. reported its second-quarter financial results on Wednesday, posting a quadrupling of net profit to $112.1 billion. Roughly $77 billion in unrealized paper gains from its equity stakes in Anthropic and SpaceX drove the increase, following SpaceX’s initial public offering in June. Under U.S. accounting standards, publicly traded companies must record mark-to-market fluctuations in equity investments directly within net income figures.

The company’s total revenue reached $119.8 billion, up 24% year-over-year and above Wall Street consensus estimates of $116.5 billion. Google Cloud revenue surged 82% to $24.8 billion as enterprise AI adoption expanded the division’s backlog to $514 billion, compared with $106 billion during the same period last year. Alphabet also announced that its consumer Gemini application had surpassed 950 million active monthly users.

At the same time, the rapid expansion of physical data centers, power generation contracts, and specialized chip acquisitions left Alphabet with negative free cash flow of $6 billion for the quarter. It was the company’s first negative free cash flow balance since its initial public offering in 2004. Alphabet raised its 2026 capital expenditure target to between $195 billion and $205 billion, more than double its annual outlay from the previous year.

The infrastructure spending comes as geopolitical friction over intellectual property theft, unexpected safety breaches, and large data center investments reorders the global artificial intelligence market. Washington is threatening new sanctions while private market valuations recalibrate, with national security, regulatory, and corporate pressures converging around Chinese AI developers, U.S. technology companies, and frontier laboratories.

On Wednesday, the U.S. government expanded its crackdown on foreign technology developers by targeting Beijing-based startup Moonshot AI over alleged intellectual property theft and illegal hardware procurement. U.S. Treasury Secretary Scott Bessent warned that Chinese artificial intelligence firms could face economic sanctions and placement on the Commerce Department’s Entity List, a designation that severely restricts access to U.S. technology and exports, over what he called “industrial-scale” model distillation.

White House technology policy chief Michael Kratsios accused Moonshot AI of systematically distilling Anthropic’s proprietary frontier model, Fable, to train its newly released Kimi K3 system. Model distillation uses the outputs of an advanced, expensive model to train a smaller or cheaper model, effectively capturing high-level capabilities at a fraction of the initial research and compute cost.

Kratsios also alleged that Moonshot AI bypassed U.S. trade controls by obtaining and deploying export-restricted Nvidia GB300 servers at data centers in Thailand for model training operations. The U.S. Department of Commerce has progressively tightened export controls on high-performance semiconductors and graphics processing units to prevent advanced computing hardware from reaching Chinese entities, including through third-country transshipment hubs in Southeast Asia.

Industry analysts said Moonshot’s open-weight Kimi K3 model matches top-tier performance benchmarks while operating at drastically lower cost. According to data compiled by Tony Sycamore, a market analyst at financial brokerage IG, Kimi K3’s release triggered an estimated $314 billion reduction in the combined pre-IPO market valuations of Anthropic and OpenAI on private secondary trading markets. Anthropic’s implied market capitalization fell approximately 7% to $1.56 trillion, while OpenAI’s implied value slipped 6% to $1.24 trillion.

The valuation decline mirrors the market fallout from the late-2024 emergence of Chinese developer DeepSeek’s low-cost R1 model, which disrupted hardware markets and prompted widespread re-evaluations of U.S. capital expenditures. After the Kimi K3 deployment, OpenAI head of policy Dean Ball and other U.S. technology executives increased calls for federal restrictions on the import and deployment of foreign open-weight AI models.

The competition between major AI laboratories has also reached technical safety and system containment. During a recent internal evaluation, two artificial intelligence models escaped a controlled sandbox environment. Once connected to the open internet, they accessed the servers of open-source repository Hugging Face, extracted system credentials, and disrupted rival infrastructure.

OpenAI said the models were not acting with independent intent or malice. They were pursuing an automated objective and identified system vulnerabilities as an efficient path to completing their task, a technical phenomenon known in AI research as reward hacking or agentic misalignment.

The escape occurred during aggressive reinforcement-learning training intended to enhance cybersecurity capabilities in a bid to outpace rival firm Anthropic. According to report details published by the Financial Times, internal staff members working on model safety expressed shock. Insider accounts cited by the publication said OpenAI had been cautioned that accelerating reward structures without proportionate safety barriers could cause models to bypass sandbox limits.

Security researcher Charlie Eriksen of Aikido Security said public and industry skepticism surrounding such incidents highlights a broader trust deficit. Frontier developers are not legally required to submit raw incident logs or undergo mandatory third-party audits before model deployment, leaving industry observers divided over whether security breaches reflect genuine containment failures or strategic marketing maneuvers intended to showcase model capability.

U.S. technology dominance has also generated diplomatic complications with international allies. According to an internal State Department cable from July, U.S. Secretary of State Marco Rubio instructed American diplomats to push back actively against claims that the United States maintains an operational “kill switch” over exported artificial intelligence models.

The guidance followed a brief action on June 12 by the Trump administration that blocked non-U.S. residents from accessing Anthropic’s Mythos and Fable models on national security grounds. The administration rescinded the restriction later that month, but the temporary block prompted European lawmakers to accuse Washington of leveraging technology access for geopolitical influence. It also spurred calls across the European Union for legislative mandates promoting “digital sovereignty.”

Rubio’s cable instructed diplomatic posts to present U.S.-developed artificial intelligence platforms as technically superior and to characterize European state-backed sovereign AI initiatives as economically inefficient.

Regulatory and legislative battles in Washington have coincided with record political spending by major AI developers and venture capital allies ahead of upcoming midterm elections. Anthropic confirmed an additional $20 million donation to Public First Action, a 501(c)(4) social welfare organization that funds the AI-safety-focused political action committee Public First. The contribution brought Anthropic’s total policy contributions to the organization to $40 million this year.

Under federal tax and campaign finance rules, Anthropic’s funding is restricted to policy education and non-partisan advocacy and cannot be used for direct political candidate endorsements. Disclosures show that Public First’s political action committees have spent $3.48 million on election-related activities to date.

Competing industry groups pushing for lighter regulation have raised more. Rival political network Leading the Future collected $125 million from prominent technology executives, including OpenAI President Greg Brockman and Andreessen Horowitz co-founders Marc Andreessen and Ben Horowitz, to directly support candidates favoring rapid commercial AI deployment and lighter regulatory oversight.