AI news for June 24, 2026

Multiple leading AI researchers are leaving Google for competitors: Jonas Adler and Alexander Pritzel are joining Anthropic, Noam Shazeer is going to OpenAI, and John Jumper (Nobel laureate for AlphaFold) is also joining Anthropic.

• Adler and Pritzel were key figures in developing Google's Gemini model.
• Noam Shazeer has been at Google since 2000 (except for his time building Character.AI, which Google acquihired for $2.7 billion).
• John Jumper won the 2024 Nobel Prize in Chemistry for his work on AlphaFold.
• The trend could accelerate as OpenAI and Anthropic prepare for IPOs and can attract top talent with equity incentives.

For a German business, this signals that top AI talent is increasingly migrating to well-funded startups — a sign of shifting industry dynamics and potential delays in Google's AI roadmap.

For you Watch how Google's AI capacity evolves; Anthropic and OpenAI could innovate faster if they deploy this talent effectively.

Micron, the largest U.S. memory chip maker, reports record profits: stock price rose from $83 (early 2024) to $1,048.51, revenue quadrupled to $41.45 billion, and profit jumped from $1.88 billion to $28.2 billion.

• Micron's market cap is now $1.2 trillion (previously ~$91 billion).
• Q3 revenue: $41.45 billion (4x year-over-year); Q4 forecast: $49–51 billion.
• Micron signed a supply deal with Anthropic for memory and storage chips and participated in Anthropic's Series H funding round.
• The AI boom has created a critical memory shortage that experts predict will persist through 2027.

The memory chip shortage is a structural business risk for AI infrastructure investments; Micron's success shows that hardware suppliers are currently capturing enormous profits while buyers face supply constraints and price increases.

For you Review your AI infrastructure roadmap for memory dependencies and negotiate early with suppliers before prices rise further.

After months of encouraging maximum AI usage, companies like Accenture are now sharply cutting back and rationing token budgets as costs spiral and value remains unclear.

• Accenture is actively preventing employees from using AI for simple tasks like PDF-to-presentation conversion.
• CFO, COO, and CIO-level leaders are questioning whether AI spending actually delivers value.
• The shift is described as a move from 'tokenmaxxing' (maximizing usage) to 'token rationing'.
• The AI selloff has hit memory chip makers particularly hard as the economic viability of AI models faces pressure.

For German SMEs, this is a reality check: AI budgets are finite, and implementation alone does not guarantee ROI. Companies must now strategically prioritize where AI creates genuine value instead of experimenting across the board.

For you Critically review your AI spending: which use cases deliver measurable cost savings or revenue growth? Focus on those instead of deploying AI everywhere.

Anthropic‑ and OpenAI‑backed groups together spent $27.41 million in the NY‑12 Democratic primary; Alex Bores narrowly lost to Micah Lasher.

• Total spending per FEC filings: $27.41 million
• Pro‑Bores super PACs (Jobs and Democracy PAC, Dream NYC, You Can Push Back, Guardrails Alliance) spent $19.26 million
• Leading the Future (partly funded by OpenAI, Palantir, Andreessen Horowitz execs) spent $8.15 million
• Final ballot count: Micah Lasher 39.1%, Alex Bores 35%

Major AI companies are channeling substantial political spending into local races over AI policy, testing the limits and effectiveness of industry‑backed political influence on regulation.

For you Monitor local and state AI legislation and lobby activity that could affect compliance costs or market access for your business.

A Snowflake benchmark finds GLM‑5.2 matches Opus 4.7 overall on 103 coding tasks but is more token‑hungry and less consistent while being cheaper per token.

• 103 tasks run three times: solved rate GLM 66% vs Opus 67%
• First‑attempt accuracy: Opus 53.7% vs GLM 47.6%
• GLM averaged 99 runs per task and used 860M tokens vs Opus's 80 runs and 439M tokens
• Pricing (from Zhipu and comparative sheets): GLM $1.40/M input, $4.40/M output; Opus $5/M input, $25/M output

Lower per‑token pricing from models like GLM‑5.2 can disrupt economics for coding use cases, but higher token consumption and lower first‑pass correctness affect latency and operational cost.

For you Benchmark alternative models on your actual dev tasks and include token usage and first‑attempt success in cost and SLA calculations before switching providers.

At Config 2026 Figma launches Code Layers, native motion and 3D/depth features, WebGPU‑powered shaders, and deeper agent/Weave integrations for designers and engineers.

• Code Layers: work with repos directly on the canvas, generate code from designs, apply changes via branches/commits/PR‑style flow; beta rollout
• Motion: built‑in animations, transitions and 3D transforms with AI prompting, timeline editing and pipeline to production via Dev Mode and MCP
• Depth & Shaders: new depth layer for perspective controls; shader effects (dither, pixelate, multiple blurs, frosted glass/chrome) implemented using WebGPU
• Agents & Weave: team‑level agent skills and richer context; planned deeper integration with Weave workflows (20+ tools) later this year

Bringing code, motion and AI into one collaborative canvas shortens the designer‑engineer loop and speeds prototyping; relying on external models shifts the value toward human judgment and workflow tooling.

For you Evaluate the Code Layers beta on a small cross‑functional project to see if it simplifies handoffs and fits your repo/CI workflow.

Google DeepMind integrates computer use natively into Gemini 3.5 Flash, enabling developers to build agents that autonomously control browsers, mobile, and desktop environments.

• Computer use was previously available only as a separate Gemini 2.5 model; now integrated directly into 3.5 Flash.
• Enables agents for long-horizon tasks and enterprise automation such as continuous software testing and knowledge work across professional applications.
• Google employs adversarial training against prompt-injection risks and offers optional enterprise safeguards for live environments.
• Available via Gemini API and Gemini Enterprise Agent Platform.

For German SMEs, this opens a new avenue to automate repetitive cross-system office processes without building custom infrastructure. It lowers the barrier to agentic automation and reduces development overhead.

For you Evaluate whether your most frequent manual workflows—data entry, cross-system processes—could be automated with Gemini 3.5 Flash; native integration may get you to production faster than custom solutions.

Hugging Face Transformers v5 can be combined with NVIDIA's NeMo AutoModel to significantly speed up and reduce GPU memory for fine‑tuning Mixture‑of‑Experts models.

• Combination: Transformers v5 (with MoE support) + NVIDIA NeMo AutoModel
• Throughput: 3.4–3.7× higher training throughput for MoE fine‑tuning vs native Transformers v5
• Memory: 29–32% less GPU memory use with the same from_pretrained() API and no code changes
• Tech: NeMo adds Expert Parallelism, DeepEP fused all‑to‑all dispatch, and TransformerEngine kernels

MoE architectures are becoming dominant for frontier models; infrastructure optimizations like DeepEP and specialized kernels lower cost and time for businesses adapting large models.

For you Evaluate whether your ML workloads benefit from MoE and benchmark NeMo AutoModel on a dev GPU setup to measure throughput and memory gains.

OpenAI reveals Jalapeño, a custom ASIC designed with Broadcom specifically for AI inference, currently in testing.

• Chip name: Jalapeño; purpose-built ASIC for inference workloads
• Developed in partnership with Broadcom; testing underway
• Early results claim substantially better performance‑per‑watt than current state‑of‑the‑art
• Targeted at real‑time models (ChatGPT/Codex); heavy pre‑training likely to remain on Nvidia GPUs

A purpose‑built inference chip can reduce OpenAI’s reliance on Nvidia and lower per‑request operating costs, affecting pricing and scalability for real‑time AI services.

For you Assess whether your cloud/AI vendor mix will change when OpenAI deploys Jalapeño; update TCO models and vendor contracts for potential shifts in inference pricing.

OpenAI and Broadcom introduce Jalapeño, a custom accelerator ASIC designed for large‑language‑model inference, targeted to run at scale by late 2026.

• Jalapeño is OpenAI’s first ‚Intelligence Processor‘, designed from scratch for modern LLM inference.
• Broadcom provides silicon manufacturing and networking tech (including Tomahawk); Celestica handles boards, racks and system integration; OpenAI leads chip design.
• Engineering samples already run ML workloads in the lab, including GPT‑5.3‑Codex‑Spark; early self‑reported tests claim substantially better performance per watt.
• The end‑to‑end design to tape‑out reportedly took nine months; Jalapeño is the first generation of a multi‑generation platform between OpenAI and Broadcom.

Owning hardware reduces reliance on third parties, can cut operating costs and latency, and gives OpenAI an edge in deploying LLM services at scale — shifting competition toward full‑stack optimization rather than model IP alone.

For you Check if your AI vendors plan custom hardware or optimize for inference accelerators; re‑evaluate long cloud commitments if providers roll out cheaper, purpose‑built inference silicon.

OpenAI grows its deployment subsidiary DeployCo by acquiring UK consultancy Tomoro (about 150 forward‑deployed engineers) and operating sites in Paris, London and Munich.

• Arnaud Fournier has been CTO of DeployCo since April 2026.
• Tomoro acquisition brings roughly 150 forward‑deployed engineers and deployment specialists into DeployCo.
• DeployCo runs sites in Paris, London and Munich and runs joint projects with German firms.
• DeployCo positions itself alongside but separate from consulting alliances such as the Frontier Alliance (Accenture, Capgemini, BCG, McKinsey).

OpenAI is moving from selling models/APIs toward embedding AI into business processes via in‑house deployment teams; that shifts where value is created and how companies will procure AI services.

For you Assess whether your digital transformation plan includes vendor‑led deployment and compliance support; consider engaging or benchmarking DeployCo when planning large LLM rollouts.

Anthropic introduces Claude Tag to embed a channel‑bound Claude in Slack; the company says an internal version now generates 65 percent of its product team's code.

• Admins assign a Claude identity to specific Slack channels; users tag @Claude to delegate tasks in natural language.
• Channel Claudes connect to selected tools, data, and codebases and build context limited to their channel purpose.
• Ambient mode lets @Claude proactively surface information, follow up on stalled threads and run asynchronous tasks over hours or days.
• Enterprise controls include per‑channel tool/data permissions, token limits and full action logs; Anthropic says Claudes won't share from private channels.

Embedding a shared AI per channel turns assistants into collaborative team members, speeding workflows like coding and ticket handling — but it also raises governance, access and audit requirements for businesses.

For you Evaluate Slack‑integrated agents for cross‑team tasks and set clear permission/audit policies before connecting code or sensitive data.

Mistral launches OCR 4, which detects document layout blocks, emits confidence scores and was preferred over competitors in 72 percent of a 600+ document blind test.

• OCR 4 extracts text from PDFs, Word and PowerPoint and classifies each element’s position and role (title, table, equation, signature).
• Provides block classification and confidence scores per word/page to aid search systems and agent pipelines.
• Supports 170 languages; independent reviewers preferred OCR 4 in 72% of blind test cases across 600+ documents.
• Available via API, Mistral Studio and Microsoft Foundry; pricing: $4 per 1,000 pages or $2 in batch mode.

Layout awareness plus confidence estimates improve downstream indexing and automated processing of documents, which matters for enterprises digitizing multilingual archives and automated workflows.

For you Run a trial of OCR 4 on a slice of your multilingual document corpus and compare layout extraction and confidence handling to your current OCR solution.

Indian customer‑engagement firm MoEngage buys San Francisco startup Aampe in an all‑cash deal to push agent‑based personalized marketing.

• Acquisition terms undisclosed; a source tells TechCrunch deal is worth tens of millions of dollars
• Aampe, founded 2020, serves 30+ customers across US, Europe and APAC and grew ARR 150% last year
• About 20 Aampe employees join MoEngage
• MoEngage raised $280M in a mix of primary and secondary transactions roughly six months earlier

Moving from segment/campaign rules to per‑customer AI agents lets brands automate who to target, what to send and when — a potential competitive edge for marketing platforms and their clients.

For you Assess whether your marketing stack can integrate agentic personalization; if you run many customer journeys, pilot agent‑based messaging to measure uplift and potential vendor migration risk.