OpenAI and Broadcom Unveil Jalapeño, Anthropic Accuses Alibaba of Distilling Claude, and Qualcomm Bets $15B on Data-Center Chips

OpenAI unveils its first AI inference chip with Broadcom, Anthropic accuses Alibaba of running a large-scale distillation campaign against Claude, Qualcomm targets $15 billion in data-center chip sales, and Google ships computer use in Gemini 3.5 Flash.

June 25, 2026 | Reading time: 8 minutes | Issue #194

Lead

OpenAI and Broadcom unveiled Jalapeño on June 24, OpenAI's first Intelligence Processor: an accelerator purpose-built for large language model inference. The chip moved from design to production tape-out in nine months, a timeline OpenAI says was accelerated by its own models. Engineering samples are already running ML workloads in the lab at production target frequency and power, including GPT‑5.3‑Codex‑Spark. Broadcom handled silicon implementation, Celestica managed board and rack integration, and Broadcom's Tomahawk networking silicon is part of the platform.

The announcement frames OpenAI's long-term strategy clearly: control the full stack from products to models to silicon. OpenAI says early testing shows Jalapeño will deliver "substantially better" performance per watt than current state-of-the-art options, and that it is designed for deployment at gigawatt scale across multiple generations. A detailed technical report is promised in the coming months.

Jalapeño is not yet a competitive threat to NVIDIA in the training market. It is aimed squarely at inference, the part of the AI workload that scales with users rather than with pre-training capex. If the performance claims hold, OpenAI can lower serving costs for its own products while reducing dependence on external supply chains. The real test is whether a single chip family can remain competitive as model architectures keep changing; OpenAI's bet is that its internal model roadmap gives it enough visibility to design silicon ahead of the curve.

Anthropic Says Alibaba Distilled Claude at Scale

Anthropic sent a letter to U.S. officials on June 24 accusing Alibaba of an "adversarial distillation" campaign against Claude. The company said Alibaba accessed Claude 28.8 million times from April to June through roughly 25,000 accounts, extracting model outputs to train its own systems. Bloomberg, Reuters, the Financial Times, CNBC, and Nikkei Asia all reported the letter, which Anthropic described as the largest known distillation effort against a U.S. AI model.

The incident exposes a structural weakness in API-based AI: model outputs are easy to collect at scale, and the line between legitimate use and systematic extraction is fuzzy until it is not. Anthropic's own business model depends on broad API access; the challenge is enforcing usage limits without undermining the developer ecosystem. Alibaba has not publicly responded to the specific claims.

The timing matters. U.S.-China AI tensions are already high over chips, cloud access, and export controls. Anthropic's complaint gives Washington a concrete example of model theft tied to a major Chinese tech company, which could feed into future trade or sanctions policy. For other AI labs, the lesson is operational: rate limits and abuse detection are now national-security infrastructure.

Qualcomm Targets Data Center Revenue of $15B by 2029

Qualcomm unveiled the Dragonfly C1000 data-center CPU on June 24, a processor built for agentic AI that CEO Cristiano Amon said Meta will use when production starts in 2028. The same day, Qualcomm raised its fiscal 2029 non-handset revenue target to $40 billion, up from $22 billion, and said it aims for $15 billion in data-center sales that year. The stock rose 15% after hours.

The announcement is Qualcomm's most aggressive data-center push since its smartphone-chip dominance. The company also disclosed a deal to acquire Modular, a startup building chip software that Qualcomm compares to NVIDIA's CUDA, for nearly $4 billion. Wired and CNBC reported the acquisition will close in the second half of 2026.

Qualcomm's pitch is power efficiency. Agentic workloads run continuously and may favor CPUs that do not draw the power budgets of the largest accelerators. The risk is timing: 2028 is far enough out that hyperscaler roadmaps could shift, and NVIDIA's ecosystem lock-in remains deep. Still, the combination of a customer like Meta, a custom-CPU strategy, and a software acquisition shows Qualcomm is serious about being a data-center player, not just a smartphone supplier.

Google Ships Computer Use in Gemini 3.5 Flash

Google made computer use a built-in tool in Gemini 3.5 Flash on June 24, available through the Gemini API and the Gemini Enterprise Agent Platform. The model can see, reason, and take action across browser, mobile, and desktop environments. Google emphasized safeguards: adversarial training against prompt injection and optional enterprise controls that require human confirmation before actions execute.

Computer use has become the next frontier after chat and coding agents. Anthropic pioneered broad deployment with Claude's computer use feature; OpenAI followed with Operator. Google's move puts the capability into a faster, cheaper model and ties it directly to its enterprise agent platform. The pitch is enterprise automation: software testing, knowledge work across professional applications, and workflow execution.

The practical challenge is reliability. Agents that can click, type, and submit are also agents that can make consequential mistakes at scale. Google's safeguards are an acknowledgment that trust, not capability, is the binding constraint for enterprise adoption.

Open-Source Pulse

Chinese open models continue to win on price, and U.S. developers are noticing. Rest of World reported this month that independent developers and startups including the San Francisco-based assistant company Lindy have switched from Anthropic models to DeepSeek, citing cost savings of millions of dollars. An Artificial Analysis index ranks DeepSeek, Xiaomi MiMo, and MiniMax among the most cost-efficient models available.

The dynamic is straightforward: Chinese labs benefit from lower domestic salaries and infrastructure costs, and many have released open-weight models or subsidized API plans to attract users. For routine tasks the quality gap is narrowing, which makes the price gap decisive. The risk for U.S. labs is not that Chinese models win on the frontier tomorrow; it is that they capture the long tail of lower-value inference that subsidizes frontier research today.

This is why Anthropic's distillation complaint is not only a security story. It is also a commercial story about who gets paid for the intelligence that flows through the global API layer.

Eastern Front

Chinese universities are reshaping curricula to prioritize AI over traditional humanities. Rest of World reported on June 22 that dozens of Chinese universities have cut foreign-language and translation programs while adding majors in "embodied intelligence" and the "low-altitude economy." A survey of 70 universities found cuts in Japanese, German, and translation studies, alongside new tech-focused degrees approved by the Ministry of Education.

The shift is more than educational fashion. It reflects a bet that embodied AI, robots, drones, and autonomous systems will drive the next wave of Chinese industrial growth. In April, the Ministry of Education approved nine universities to begin enrolling students in embodied intelligence, and 38 new majors overall for the upcoming academic year, most of them tech or digitalization-related.

The contrast with U.S. higher education is instructive. American universities are still debating AI policy in classrooms; Chinese universities are redesigning degree programs around it. The long-term effect will show up not in next quarter's model release but in the composition of the engineering workforce five years from now.

India Lens

India's strategy for AI sovereignty is taking shape through partnerships that keep hardware on Indian soil without relying solely on U.S. clouds. Rest of World reported on June 1 that G42, backed by Abu Dhabi's Mubadala, signed an agreement in May to deploy an AI supercomputer in India built from 64 Cerebras systems. A G42 unit will handle installation, operations, and maintenance; data will remain under Indian governance rules.

The deal is a second path for a country that has collected more than $45 billion in cloud commitments from Amazon, Microsoft, and Google. India's national AI program already makes 34,000 Nvidia GPUs available to researchers and businesses, with a target of 100,000 by year-end. The G42-Cerebras arrangement adds a non-U.S. option: sovereign compute, operated locally, financed by Abu Dhabi.

For India, the logic is pragmatic. Full AI sovereignty is impossible because no country controls the full stack; assembling capabilities from multiple partners is the realistic alternative. The risk is execution. Commitments are rising, but power, land, and chip availability remain binding constraints on the ground.

Europe

The European Commission published guidelines on prohibited AI practices under the AI Act on June 24, clarifying rules that took effect in February 2025. The guidelines cover harmful manipulation, social scoring, untargeted scraping for facial recognition databases, emotion recognition in workplaces and schools, and real-time remote biometric identification for law enforcement. The Commission says the guidelines are designed to ensure consistent application across the EU, though they are non-binding and authoritative interpretation remains with the Court of Justice.

The publication is part of a broader implementation push. High-risk system obligations arrive in 2027, and the Commission is trying to make the AI Act's prohibitions concrete enough to enforce before the heavier compliance phases begin. For AI companies, the guidelines offer more clarity but no safe harbor. The real test will be the first enforcement actions, which will establish how aggressively the EU intends to interpret its new authority.

The View

Four announcements on June 24 point to the same underlying tension: AI infrastructure is becoming a sovereignty contest, not just a technology market. OpenAI is building its own chips. Anthropic is asking governments to police model extraction. Qualcomm is selling data-center silicon to Meta with an eye on NVIDIA's margins. And the EU is writing the rules that will govern how all of it can be deployed.

The common thread is control. Labs want control over supply chains, outputs, and distribution. Governments want control over data, models, and deployment. Users, meanwhile, are choosing cheaper models wherever they come from, which is why Chinese open models are gaining traction among U.S. developers.

The market is splitting into layers. Frontier training remains concentrated among a few well-capitalized labs. Inference is fragmenting across chips, geographies, and price points. And regulation is fragmenting across jurisdictions. The winners will be the companies and countries that can operate across all three layers without being trapped in any single one.

The Miss

NVIDIA launched the BioNeMo Agent Toolkit for scientific discovery on June 23, giving AI agents domain-specific tools for protein structure prediction, molecular docking, generative chemistry, and genomic analysis. The toolkit is being adopted by Dassault Systèmes, Databricks, Eli Lilly, Schrödinger, Snowflake, and the UW Medicine Institute for Protein Design; Anthropic and OpenAI are integrating it.

The announcement was buried under chip-stock volatility and product launches from the big labs. It matters because scientific discovery is the use case most likely to justify AI's infrastructure spending in the long run. BioNeMo is an early platform layer for agentic science; if it works, the economic returns will exceed those of content-generation tools by orders of magnitude.

Pull Quotes

"The world is moving to a compute-powered economy. Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant." — Greg Brockman, OpenAI

"Frontier models are the brains. BioNeMo is the scientific toolbox. Together, they give AI agents the skills of a PhD research assistant and the speed of a supercomputer." — Jensen Huang, NVIDIA

"You don't need God to write your email. If you can get those lower tiers of intelligence for a tenth of the price, it would be foolish not to do it." — Flo Crivello, Lindy, on switching to DeepSeek

"This is an example of India's pragmatic approach to AI sovereignty, using the power of its scale." — Cameron Kerry, Brookings Institution

Reads & Links

OpenAI and Broadcom unveil Jalapeño — OpenAI on its first inference chip, designed in nine months with Broadcom and Celestica. https://openai.com/index/openai-broadcom-jalapeno-inference-chip/

Anthropic accuses Alibaba of illicitly accessing Claude — Bloomberg on the distillation complaint and 28.8 million API calls. https://www.bloomberg.com/news/articles/2026-06-24/anthropic-accuses-alibaba-of-illicitly-accessing-its-ai-models

Qualcomm stock pops on $15B data-center target — CNBC on Dragonfly C1000, the Meta deal, and the $40 billion non-handset revenue forecast. https://www.cnbc.com/2026/06/24/qualcomm-data-center-cpu-meta.html

Google adds computer use to Gemini 3.5 Flash — Google blog on the built-in tool for browser, mobile, and desktop automation. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-computer-use-gemini-3-5-flash/

Chinese universities cut humanities for AI majors — Rest of World on the shift toward embodied intelligence and tech-focused degrees. https://restofworld.org/2026/chinese-universities-drop-humanities-ai/

India's AI deal with the UAE challenges U.S. cloud dominance — Rest of World on the G42-Cerebras supercomputer and Indian sovereignty strategy. https://restofworld.org/2026/india-uae-g42-cerebras-ai-sovereignty/

EU Commission publishes AI Act prohibited practices guidelines — European Commission guidelines on banned AI uses under the AI Act. https://digital-strategy.ec.europa.eu/en/library/commission-publishes-guidelines-prohibited-artificial-intelligence-ai-practices-defined-ai-act

NVIDIA BioNeMo Agent Toolkit — NVIDIA on domain-specific agent tools for scientific discovery. https://nvidianews.nvidia.com/news/nvidia-launches-bionemo-agent-toolkit-giving-ai-agents-the-tools-to-accelerate-scientific-discovery

The AI stack is splitting into sovereign layers, and June 24 showed every major player choosing which layer to own.