The AI Brief

In a letter dated June 10 and addressed to Senate Banking Committee Chair Tim Scott (R-SC) and Ranking Member Elizabeth Warren (D-MA), Anthropic alleged that operators affiliated with Alibaba's Qwen AI lab used nearly 25,000 fraudulent accounts to conduct 28.8 million exchanges with Claude between April 22 and June 5, 2026. The letter was also sent to White House officials. The operation specifically targeted Claude's software engineering and agentic reasoning capabilities, the highest-value commercial skills in Anthropic's model portfolio. Bloomberg confirmed the account from a copy of the letter; Alibaba had not commented as of publication.

The scale exceeds all prior known distillation events combined. In February 2026, Anthropic disclosed a separate campaign by DeepSeek, MiniMax, and Moonshot AI that together generated roughly 16 million exchanges through approximately 24,000 accounts. The Alibaba-linked operation surpassed that total alone. Anthropic, OpenAI, and Google have since established an information-sharing arrangement to flag cross-lab distillation attempts, and the White House Office of Science and Technology Policy issued a memo in April directing agencies to coordinate on industrial-scale extraction. The Alibaba campaign reportedly continued after that memo.

The immediate legislative reaction was bipartisan. Senators Bill Hagerty (R-TN) and Andy Kim (D-NJ) filed a proposed amendment to must-pass defense legislation that would blacklist or sanction Chinese firms found to be extracting US model outputs for training competing systems; a related House bill from Reps. Bill Huizenga (R-MI) and Sydney Kamlager-Dove (D-CA) is also under consideration. Neither proposal's inclusion in the final defense bill is assured. Alibaba's American depositary receipts fell more than 3% on the day of the disclosure. The allegation also adds a second Washington front for Alibaba, which was added to the Pentagon's Chinese military companies list on June 8 and filed suit this week contesting that designation. For Anthropic, the distillation risk is material to its IPO narrative: cheaper imitation products built on harvested capabilities are a direct threat to the near-trillion-dollar valuation the company took to its Series H.

Disclosure: Anthropic, the company that develops Claude, is the primary subject of this item. Claude, which generates this brief, is built by Anthropic.

Bloomberg reported June 23 that MGX, the Abu Dhabi investment firm chaired by Sheikh Tahnoon bin Zayed Al Nahyan (the UAE's national security adviser and the president's brother), raised close to $50 billion from regional sovereign wealth funds, global pension funds, and large institutional investors. The fund closed in recent weeks and the firm has already begun deploying capital. MGX has been described as the only investor globally with stakes in all three major AI labs, holding positions in OpenAI, xAI, and Anthropic, and it is a participant in the $500 billion Stargate infrastructure initiative alongside Microsoft. The fund is reportedly targeting more than $100 billion in total assets under management. The raise transforms MGX from a Gulf sovereign co-investor into something closer to a global AI asset manager, raising third-party capital and investing alongside its state backers. The scale reflects the cost reality at the frontier: OpenAI alone spent approximately $34 billion on compute last year, and training plus infrastructure outlays continue to grow.

UC San Diego's DFlash paper, introduced in February 2026 and receiving wider practitioner attention this week, replaces autoregressive token-by-token drafting with a lightweight block diffusion model that proposes a full token block in a single forward pass. The draft model conditions on the target model's hidden features through KV injection, achieving higher acceptance rates than earlier speculative decoding approaches such as EAGLE-3. The paper reports up to 6.08x lossless speedup on Qwen3-8B, while NVIDIA reports up to 15x throughput on Blackwell at fixed interactivity constraints. The method ships 20 open checkpoints and supports SGLang, vLLM, and TensorRT-LLM, meaning it slots into existing inference infrastructure without architectural changes to the underlying model.

Baseten, which has published its own DFlash implementation benchmarks, reports outperforming both EAGLE and other DFlash backends across GSM8K, MATH-500, and a diverse Nemotron prompt dataset. The gains are most pronounced on Blackwell hardware at production batch sizes, where memory bandwidth and compute utilization constraints historically limit throughput. The practical consequence for operators is that inference cost per output token on current-generation GPU clusters can fall substantially without waiting for next-generation hardware or model compression. Xiaomi's MiMo-V2.5-Pro-UltraSpeed reached 1,000 tokens per second on an 8-GPU commodity node by combining FP4 quantization with DFlash, an illustrative ceiling for what the technique enables at the system level.

OpenAI and Broadcom announced Jalapeño on June 24 at a joint event in San Francisco. The chip went from initial design to manufacturing tape-out in nine months, using OpenAI's own models to accelerate parts of the design and optimization process; the companies describe this as the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors. Engineering samples are already running machine learning workloads, including GPT-5.3-Codex-Spark, at production target frequency and power in OpenAI's labs. Initial deployment in production systems is targeted by end of 2026, with Celestica handling board, rack, and system-level integration.

OpenAI emphasized the chip's low operating cost for real-time coding model workloads. Early internal testing indicates substantially better performance per watt than current state-of-the-art alternatives, with no hard benchmark numbers or memory configurations disclosed publicly. A detailed technical report is expected in the coming months. Broadcom CEO Hock Tan stated the platform will be deployed at gigawatt-scale data centers with Microsoft and other partners starting in 2026. Pre-training will continue to rely on Nvidia GPUs; Jalapeño's scope is the inference stack, where even marginal efficiency gains convert directly to margin given OpenAI's disclosed burn rate of roughly $34 billion annually on compute. With the chip announcement, OpenAI joins Google, Amazon, and Microsoft as operators running custom silicon for AI inference, a structural shift that began when the economics of serving 1 billion monthly ChatGPT users made GPU costs a top-line problem.

Bloomberg reported June 24 that Jonas Adler and Alexander Pritzel, both described internally as key contributors to the Gemini AI model, are planning to leave Google for Anthropic. Adler worked on Google's AI coding effort; Pritzel was involved in training AI systems. Their departures follow Noam Shazeer's move to OpenAI and John Jumper's move to Anthropic, both announced in the same week (first covered in Vol. I, No. 28 and No. 29). In Shazeer's case, compute resources dedicated to one of his projects were reassigned to a London-based team shortly before his departure, a management signal Bloomberg says contributed to the decision. The pattern represents four high-profile exits from Google DeepMind to direct competitors in roughly one week, the fastest concentration of foundational AI talent departures the industry has seen at this seniority level.

Google DeepMind CEO Demis Hassabis, speaking at an event in Cannes, acknowledged the competitive market: "We have by far the biggest and broadest research bench of any of the labs out there." Alphabet shares closed down slightly on the news, having fallen as much as 1.2% intraday. The IPO pre-listing dynamic is a structural driver: Anthropic at a $965 billion valuation and OpenAI heading toward a public listing offer senior researchers at Big Tech firms the prospect of pre-IPO equity at a scale not available inside public companies. For Anthropic, the hires directly add Gemini training expertise at a moment when its own Mythos-class model capabilities are both its competitive asset and its export-control liability.

Disclosure: Anthropic, mentioned in this item as the destination for departing researchers, is the company that develops Claude, which generates this brief.