The AI Brief

Bloomberg first reported, and CNBC confirmed, that bankers leading the Anthropic IPO have begun scheduling meetings between prospective institutional investors and Anthropic executives in the coming weeks. The company filed its IPO prospectus with the SEC confidentially last month and is targeting a debut as soon as October, though timing remains subject to change. Goldman Sachs, Morgan Stanley, and JPMorgan Chase are the lead underwriters.

Anthropic enters the process from a position most IPO candidates lack. The company closed a $65B funding round in May at a $965B valuation, eclipsing OpenAI's $852B valuation for the first time, and carries an implied valuation of roughly $1.2T on secondary markets. Annual revenue is reported at a $47B run rate. OpenAI, which had previously targeted a fall 2026 listing, has pushed its debut to 2027, leaving Anthropic positioned to be the first frontier lab on public markets. DeepSeek is also pursuing an IPO but on a later timeline.

The path is not unencumbered. Anthropic previously faced a US government export control order suspending two of its models for 18 days, sued the Defense Department after it was classified as a supply-chain threat, and operates under a Long-Term Benefit Trust structure with constraints on profit distribution that public market investors will need to evaluate. The October target lands during enterprise budget season, when CIOs are making 2027 AI commitments, a calendar alignment Anthropic's bankers are likely treating as a feature.

Disclosure: Claude, which generates this brief, is built by Anthropic.

TSMC reported Q2 2026 results on July 16, with revenue reaching $40.2B, up 33.7% year-over-year in dollar terms and 36% in NT dollar terms, hitting the upper end of its own guidance range. Net profit climbed 77.4% to a record NT$706.56 billion, significantly exceeding analyst expectations. Gross margin reached 67.7%, above the guided 65.5%–67.5% band, attributed to high fab utilization and disciplined cost management despite ongoing overseas fab ramp-up costs.

The structural shift is the HPC (High Performance Computing) share: AI-related demand now constitutes 66% of total TSMC revenue, up from approximately 52% a year earlier. CoWoS (Chip-on-Wafer-on-Substrate) advanced packaging remains fully booked through year-end. TSMC raised its 2026 capital expenditure outlook toward the upper end of its $52B–$56B range, signaling continued capacity commitment rather than a demand-driven pull-back.

First covered in Vol. I, No. 50, where Q2 revenue of $39.6B was the reported figure based on monthly disclosures. Today's full earnings call adds net profit, margin performance, and forward guidance, the numbers that move investment theses. Management emphasized that demand from agentic AI workloads, which require more compute per query than generative AI alone, is lengthening the demand curve beyond what analysts modeled at the start of the year.

The Future of Life Institute published its Summer 2026 AI Safety Index, grading nine major labs across six domains: risk assessment, current harms, safety frameworks, existential safety, governance and accountability, and transparency. An independent panel of seven AI researchers and governance experts, affiliated with institutions including UC Berkeley, University of Montreal, and Oxford, reviewed public model cards, research papers, benchmark results, and survey responses. Five of nine companies completed the survey; Alibaba, xAI, DeepSeek, and Mistral did not respond.

Anthropic scored highest at C+, followed by OpenAI and Google DeepMind at C, Meta at D+, Z.ai (Zhipu) and Alibaba Cloud at D-, and xAI, DeepSeek, and Mistral at F. The three failing labs represent one each from the US, China, and Europe, the panel's way of underscoring that inadequate safety is not a geography problem. xAI's score collapsed from a stable D throughout 2025 to 0.65 this cycle. The index flagged that Anthropic, OpenAI, Google DeepMind, and Meta have each weakened or eliminated earlier commitments to pause development if systems approached specified danger thresholds, what the panel calls "red line" commitments. Military AI use is the sharpest flashpoint: labs that banned defense applications between 2024 and 2026 have reversed course and sought defense partnerships.

The index covers evidence through June 3, 2026. Mistral pushed back in comments to AFP, arguing the evaluation framework is poorly suited to open-source development. The FLI panel acknowledged the open-weight complication but noted Mistral declined to participate in the survey process. OpenAI leads the Risk Assessment subcategory specifically, on the strength of its external evaluation suite and METR engagement. The report's overall conclusion: voluntary self-policing has begun eroding before durable regulatory alternatives are in place.

As of July 16, Google's official Gemini API documentation lists gemini-3.5-flash and gemini-3.1-pro-preview, no gemini-3.5-pro model ID. The July 17 date circulating in media coverage comes from leaks and third-party reporting, not an official Google announcement. A fallback date of July 24 has also been reported. Sundar Pichai first promised the model at Google I/O on May 19, "give us until next month", which drew audible groans from developers. June came and went. A mid-July target set after a full base-model rebuild has now also approached without a signed launch post. This is covered as a continuing story; first covered in Vol. I, No. 49 (July 13); slug gemini-35-pro-architecture-rebuild-july17 has appeared across four editions.

The delay's cause is documented: Google DeepMind scrapped the original model after engineers found structural failures in recursive tool-calling environments, the multi-step chains where an agent calls one tool, uses the result to call another, and so on. Recursive tool-call stability is the defining requirement for an agentic coding model, which is the stated purpose of the 3.5 generation. The rebuilt model carries leaked specs of a 2M-token context window and a Deep Think reasoning mode, both unconfirmed in official documentation.

The competitive context has hardened significantly since the June slip. GPT-5.6 Sol launched publicly on July 9; Grok 4.5 launched the same week; DeepSeek V4 Pro is in general availability. Google has meaningful distribution advantages, Search, Workspace, Android, Vertex, but distribution does not answer an API call. Gemini 3.5 Flash has been carrying production workloads since May 19 and outperforms the older Gemini 3.1 Pro on Terminal-Bench 2.1 (76.2% vs. 70.3%) at $1.50 per million input tokens.

South Korea announced a decade-long national plan valued at approximately 1,350 trillion won, roughly $880B at current exchange rates. The plan allocates approximately $518B for memory chip manufacturing capacity, approximately $550B for AI data centers, targets 8.4 gigawatts of data-center power capacity by 2029, and includes a push to grow South Korea's humanoid robotics market share from 1% to 20% by 2028. The timeline and dollar breakdown are per reporting from Build Fast With AI's July 15 coverage and Unrot's independent summary of the same announcement.

The plan comes from a government that already anchors the AI hardware stack through SK Hynix, which raised $26.5B in its Nasdaq ADR debut earlier this month on its 60% share of the AI HBM memory market, and Samsung Foundry, which competes with TSMC on advanced node fabrication. A nationally funded capacity expansion at this scale, if executed, would materially affect HBM pricing and availability through the decade, and would reduce South Korea's dependence on any single export market for its chip revenue.

No binding legislative text has been confirmed at time of publication; the announcement is a government plan rather than a signed statute or formal budget appropriation. The robotics market share target of 20% by 2028 represents the most aggressive component and the one with the least current baseline, South Korea held roughly 1% of the humanoid robotics market at the plan's announcement. The data-center and memory targets rest on existing industrial capacity that is already scaling.

Caveat Dollar equivalents are subject to won-dollar exchange rate movement; the 1,350 trillion won figure is the authoritative number. Independent verification of the final plan text is pending.