AI Business · 01 Jul 2026

AI Business | Jul 1, 2026

Analysis

The Commerce Department's reversal on Anthropic's Fable 5 landed the same day Anthropic shipped both Claude Sonnet 5 and Claude Science, pairing a resolved export-control dispute with a two-product push that leans explicitly on "lower cyber risk" positioning — the same axis that triggered the original ban — while OpenAI's GPT-5.6 remains capped near two dozen approved partners pending its own wider release.
Two of the day's biggest capital commitments target the same supply constraint from opposite ends of the value chain: South Korea's $590 billion (Won911 trillion) chip and AI infrastructure program adds memory and fab capacity, while Digital Realty's $7.8 billion Virginia transaction adds finished, leased data-center space in the world's largest data-center market — both explicitly framed around AI demand outrunning available supply.
China's AI-infrastructure story split two ways on the same day: CXMT's $2.94 billion DRAM supply deal with Tencent shows a domestic chipmaker still courting hyperscaler business ahead of its own listing, while Meituan's LongCat-2.0 — trained entirely on more than 50,000 homegrown ASICs and edging out GPT-5.5 on a coding benchmark — argues China no longer needs Nvidia GPUs to field a near-frontier open-weight model, cutting against the premise behind US export restrictions on both chips and models.
Anthropic's public call for a verifiable, industry-wide pause on frontier development ran into legal analysis the same week concluding such coordination could be treated as an illegal per se restraint of trade under the Sherman Act, and into a parallel argument that a genuine US-China AI safety deal is unlikely given the verification problem and China's own capability gains via GLM 5.2 — leaving Anthropic's near-term regulatory fix bilateral and narrow, not evidence of broader coordination.

POLICY & REGULATION

TRADEFinancial Times | White House lifts ban on Anthropic models | The US Department of Commerce lifted export controls on Anthropic's Fable 5 model Tuesday evening, ending a three-week impasse between the $1 trillion company and the Trump administration and clearing Fable 5 for public re-release, people with knowledge of the matter told the Financial Times. Commerce Secretary Howard Lutnick had imposed the ban on June 12 after the government discovered a "jailbreak" bypassing the model's safety guardrails; the government separately allowed Anthropic to release the more restricted Mythos 5 model to about 100 pre-vetted partners the prior week. Reuters reported the reversal was expected as early as Tuesday, citing a source familiar with the matter. OpenAI's GPT-5.6, similarly limited to roughly two dozen government-approved partners, could see wider release "as early as next week," per a person with knowledge of the matter.

PARTNERSHIPS & ENTERPRISE

ENTCNBC | Salesforce is on an AI buying spree, but Wall Street still has its doubts | Salesforce has pursued an acquisition run to build out its AI capability, most recently a $3.6 billion purchase of AI customer-service platform Fin, whose lead product resolves complex customer queries across email, WhatsApp, Slack, and live chat. Salesforce said Fin will complement its flagship Agentforce suite, targeting small and midsize businesses seeking rapid deployment. CNBC reported Wall Street remains unconvinced the strategy has made Salesforce an AI winner rather than a laggard.
PRODThe Verge | OpenAI is teasing new hardware... for Codex | OpenAI teased a Codex-branded hardware device built in partnership with mechanical-keyboard maker Work Louder, set to launch July 15, distinct from the AI hardware project OpenAI is separately developing with former Apple designer Jony Ive. The device's silhouette resembles Work Louder's Creator Micro 2 macro pad, which lets users assign custom shortcuts and actions across apps.

GOVERNANCE & SAFETY

GOVBloomberg | Trump Convert Marc Andreessen Gets Spot on Pentagon Policy Board | Venture capitalist Marc Andreessen, a former Democratic backer turned prominent Trump supporter and vocal opponent of AI regulation, was named alongside several other Republican politicians and former officials to the Pentagon's Defense Policy Board, a top advisory panel, Bloomberg reported.
GOVLawfare | Can Frontier AI Labs Lawfully Agree to Pause? | The Anthropic Institute published an essay arguing it "would be good for the world" to have the option for frontier labs to slow or pause development if peer companies verifiably did the same, citing METR research that the length of tasks AI systems can complete autonomously doubles roughly every four months. Legal analysis in Lawfare found that a formal, verifiable pause agreement among competing labs risks per se treatment as an illegal output restriction under Section 1 of the Sherman Act, meaning current US antitrust law may bar the very coordination Anthropic is proposing; the Department of Justice and Federal Trade Commission have opened a joint inquiry into competitor-collaboration guidance that could address the gap.
GOVFinancial Times | What would multilateral 'AI arms control' look like? | A Financial Times analysis by the author of "Chip War" questioned whether a US-China AI-safety deal is achievable given that Z.ai's newly released GLM 5.2 model prompted David Sacks, co-chair of Trump's Council of Advisors on Science and Technology, to say China now has "a Chinese open-weight model that is as good as the currently available models from OpenAI and Anthropic." The piece noted the US and China discussed AI safety ahead of the May Xi-Trump summit without agreeing to formal negotiations, and argued that verification remains the core obstacle since neither side will grant the other access to sensitive source code.

RESEARCH & MODELS

MODELAxios | Anthropic debuts Claude Sonnet 5 for everyday agent tasks with lower cyber risk | Anthropic launched Claude Sonnet 5, a more powerful and agentic version of its midsize model, which the company says approaches the performance of Opus 4.8 while posing "much lower ability" to perform dangerous cyber activity than its Opus and Mythos models. Anthropic said the release came amid ongoing discussions with the Trump administration about its models and arrives while the company still awaits full government approval to restore access to Mythos and Fable; Lovable co-founder Fabian Hedin said Sonnet 5 "refuses unsafe requests cleanly and consistently." TechCrunch noted the launch mirrors recent agentic-focused releases from OpenAI's GPT-5.6 Sol and Google's Gemini 3.5 Flash.
SCIFinancial Times | Anthropic launches Claude Science in push for pharma revenue | Anthropic launched Claude Science, its first product dedicated to scientists and pharmaceutical companies, with use cases including rendering 3D protein structures and drug discovery, as the $900 billion company seeks new enterprise revenue ahead of a planned IPO. Life-sciences head Eric Kauderer-Abrams said Anthropic will also use the product to pursue its own research into treatments for "neglected" diseases that traditional biopharmaceutical companies find commercially unattractive. TechCrunch reported the launch is "not a new AI model" but a dedicated workbench running existing Claude models, including Opus 4.8, building on Anthropic's October 2025 "Claude for Life Sciences" plug-ins; Anthropic is also offering up to 50 postdoctoral and graduate research grants worth $30,000 in credits each.
SCIBloomberg | Google Cloud to Offer Specialist AI Models for Science Research | Google will begin offering specialist AI models from SandboxAQ through Google Cloud, expanding enterprise and research access to technology designed to accelerate drug discovery, materials science, and semiconductor manufacturing.

COMPUTING & INFRASTRUCTURE

CHIPFinancial Times | Samsung and SK Hynix plan $590bn chipmaking expansion | Samsung Electronics and SK Hynix, together holding roughly 80% of the global high-bandwidth-memory market, will invest a combined Won911 trillion ($590 billion) alongside the South Korean government to build chipmaking capacity in the country's underdeveloped southwest, part of President Lee Jae Myung's "Three Mega Projects for the Great Leap Forward" spanning semiconductors, AI data centers, and robotics. The package includes Won800 trillion for four new fabs and Won81 trillion for a chip-packaging cluster; Gwangju and South Jeolla province will separately commit Won520 trillion ($337 billion), and Reuters reported local media put the full program above $650 billion. Samsung and SK Hynix now carry a combined market capitalization of roughly $2 trillion; chip-equipment makers including ASML, which closed up 6.8% in Amsterdam to a record, rallied on the announcement even as memory shortages have already pushed Apple to raise MacBook and iPad prices.
CHIPReuters | China's CXMT wins $3 billion memory supply deal with Tencent, sources say | Chinese memory chipmaker ChangXin Memory Technologies (CXMT) signed a long-term DRAM supply agreement worth more than 20 billion yuan ($2.94 billion) with Tencent Holdings, ahead of a Shanghai listing that aims to raise 29.5 billion yuan, people familiar with the matter told Reuters. CXMT, the world's fourth-largest DRAM maker with roughly 7.7% market share, posted first-quarter revenue up 700% year-on-year as it doubles wafer output toward 600,000 wafers a month; the company is also in talks on similar supply deals with Alibaba, ByteDance, Lenovo, and Xiaomi.
CHIPVentureBeat | Meituan open sources LongCat-2.0, the 1.6T, near-frontier agentic coding model that's been leading OpenRouter — trained entirely on Chinese chips | Meituan open-sourced LongCat-2.0, a 1.6 trillion-parameter Mixture of Experts (MoE) coding model under an MIT license, revealing it as the model behind "Owl Alpha," which had topped OpenRouter's developer usage charts. The model was trained entirely on more than 50,000 domestic Chinese ASICs rather than Nvidia GPUs and scores 59.5 on SWE-bench Pro, edging OpenAI's GPT-5.5 at 58.6, arriving as Washington restricts access to GPT-5.6 and Anthropic's Fable/Mythos models.
DCReuters | Digital Realty to buy Blackstone's data center stake in deal valuing assets at $7.8 billion | Digital Realty agreed to acquire Blackstone-managed funds' blended 64% stake in three fully leased Northern Virginia data centers — the world's largest data-center market — in a deal valuing the assets at $7.8 billion, paying $3.5 billion in $1.2 billion cash and $2.3 billion in shares. The three facilities, two 96-megawatt sites in Manassas and one in Sterling, were expected to close June 30; Digital Realty shares fell as much as 4% on the announcement.

GENERATIVE MEDIA & EMBODIED AI

GENArs Technica | Google's new Nano Banana 2 Lite image model is its fastest and cheapest yet | Google DeepMind released Nano Banana 2 Lite — technically Gemini 3.1 Flash-Lite Image — an image-generation model that produces images in about 4 seconds versus roughly 20 seconds for standard Nano Banana 2, at API pricing of about $0.034 per 1,000 images, roughly half the standard rate. Google said the model trades some quality for speed, struggling more with small text and infographic accuracy, and applies SynthID watermarks to all outputs; it is available now in the Gemini API, Google AI Studio, and the Gemini app.

Calendar

PRODJul 15 — OpenAI's Work Louder co-branded Codex hardware device is set to launch.
SCIJul 15 — Applications close for Anthropic's Claude Science postdoctoral/graduate research-grant program (up to 50 projects, $30,000 in credits each).
SCIJul 31 — Award notifications go out for Claude Science research grants; funded projects run Sep 1–Dec 1, 2026.

MARKETS

30 Jun 2026 close | Retrieved 30 Jun 22:33 UTC | Yahoo Finance

AI Infrastructure (1D) | CoreWeave 99.54 USD +4.2%

Indices (1D) | NASDAQ 26,214 +1.5% | SOX 14,247 +3.9%

Coverage: 29 Jun 01:00 – 1 Jul 01:00 UTC

§ Subscribe

Read AI Business every morning.

One click. Free during preview. Reply to refine or cancel.

Want a Beat tuned to your specific interests? Start here →