2026.06.24

The Full-Stack AI Arms Race — What Jalapeño and Kioxia Reveal About the New Order

The AI world packed an unusual density of infrastructure power plays into a single day. OpenAI unveiled its proprietary inference chip “Jalapeño,” Fujitsu published a new LLM architecture claiming 475× the throughput of Transformers, and Kioxia delivered a valuation surge that briefly made it Japan’s most valuable public company — a vivid reminder of where AI profits are actually accumulating. The push to own every layer — model, chip, and data — is nearing a critical inflection point.

OpenAI and Broadcom Unveil “Jalapeño” — From Design to Mass Production in Nine Months

OpenAI has announced its first internally designed AI chip, “Jalapeño,” formally named the “Intelligence Processor,” co-developed with Broadcom (NASDAQ: AVGO) and EMS leader Celestica. The standout detail is speed: the chip went from design to production in roughly nine months. OpenAI’s claim that its own models accelerated the development process hints at a self-reinforcing loop — AI designing the chips that run AI.

On performance, OpenAI describes the chip as delivering “substantially better performance per watt compared to current state-of-the-art,” though specific figures have not been disclosed. The ceremony in which Broadcom CEO Hock Tan personally handed the first chips to Sam Altman and Greg Brockman signals more than a technical milestone. Broadcom’s track record manufacturing Google’s TPUs and Meta’s AI chips means OpenAI has secured one of the world’s top partners in AI-specific silicon design.

Deployment ambitions are equally bold. OpenAI plans to roll out the chip at “gigawatt scale” with data center partners across multiple generations, explicitly targeting not just current LLMs but future architectures as well — signaling a general-purpose inference platform rather than one optimized for a single model. Reducing dependence on NVIDIA and closing the loop on model, service, and chip is a path Google and Amazon have walked before. The threat to competitors lies in OpenAI’s apparent ability to do it in nine months.

Source: OpenAI Blog - OpenAI and Broadcom unveil LLM-optimized inference chip

Fujitsu’s “PHOTON” — What’s Behind the 475× Throughput Claim

Fujitsu has announced a new LLM architecture called “PHOTON” (Parallel Hierarchical Operation for TOp-down Networks), claiming up to 475× the per-GPU throughput of Transformers. The headline number is attention-grabbing, but the details deserve careful reading.

PHOTON targets a well-known bottleneck: the KV-cache memory pressure that builds up during long-context inference and high-concurrency serving. Standard Transformers compute relationships across every token, causing memory access to scale explosively with context length. PHOTON addresses this by processing text hierarchically in semantic chunks, reducing the computation required, and further bundles multiple user queries together via “multi-query integration” to handle them in a single pass.

Validation was conducted on models at 600M, 900M, and 1.2B parameters. The 1.2B model achieved 475× multi-query throughput at the cost of “slight quality degradation” — and the paper notes that bundling just nine queries already recovers parity with a standard Transformer, suggesting the trade-off is workable in practice. Smaller KV-caches also mean more concurrent generations can fit within the same GPU memory, with direct implications for inference infrastructure cost.

The architecture is positioned for low-cost multi-agent processing, which aligns neatly with the current inference-time scaling trend. The work is accepted as an oral presentation at ACL 2026 (July 2, San Diego) — peer review adds meaningful credibility. Japanese AI research has been gaining ground at top-tier venues, but a claim of this magnitude at the LLM architecture level is unusual.

Source: ITmedia AI+ - 「Transformerの最大475倍」　富士通、GPUを効率的に使うLLMアーキテクチャ「PHOTON」開発

The Kioxia Phenomenon — A 70× Stock Rally and a Market Cap That Briefly Surpassed Toyota

A 70× stock gain in eighteen months since IPO. A market cap that briefly crossed $400 billion, overtaking Toyota as Japan’s most valuable publicly traded company. The numbers invite skepticism — but the fundamentals largely hold up under scrutiny.

Kioxia’s Q2 FY2026 (April–June) revenue is estimated at roughly $11.7 billion, with an operating margin of approximately 74%. That single quarter is on track to exceed the company’s full-year results for FY2025. CFO Yoshihiko Kawamura has characterized FY2026 as the entry point into a “super cycle,” suggesting that the structural demand from AI has elevated the memory industry’s typical supply-demand cycle to unprecedented heights.

The historical context matters. Kioxia traces its origins to Toshiba’s invention of NAND flash memory in 1987. It was spun off and sold during Toshiba’s corporate crisis and debuted in December 2024 at a price below its IPO offer — a humiliating start. The reversal since then tracks the surge in AI data center storage demand, particularly from inference workloads where model weights and large caches increasingly reside in storage tiers. HBM (High Bandwidth Memory) was initially seen as the primary AI beneficiary, but the spread of inference-time scaling has forced a reassessment of storage’s strategic importance.

For anyone tracking AI infrastructure, Kioxia is a case study confirming that demand is spiking simultaneously at Layer 1 (power and cooling), Layer 2 (compute), and Layer 3 (memory and storage). As competition intensifies at the service layer, structural profit concentration in infrastructure-layer players seems likely to persist.

Source: ITmedia AI+ - キオクシアなぜ急成長？　半導体メモリって何？　AIブームを見通すための基礎知識

Anthropic Launches “Claude Tag” for Slack — Backed by an Internal Track Record of 65% AI-Written Code

Anthropic began beta access to “Claude Tag” for Slack on June 23. The mechanic is straightforward: a dedicated Claude instance is assigned to a channel and takes on tasks when mentioned. But two capabilities set it apart from conventional Slack bots.

First, channel-scoped persistent memory. Claude Tag continuously learns from in-channel conversations, accumulating team-specific context — internal rules, interpersonal dynamics, project history — so users never need to re-explain the same background twice and Claude can pick up where a previous team member left off. Second, asynchronous parallel work: the system can “work autonomously on tasks over hours to days in parallel with other work,” meaning it can write code or update documentation in the background without blocking the conversation.

Anthropic frames this as “the beginning of Claude Code’s evolution,” and backs it with a concrete internal data point: 65% of code written by Anthropic’s own product teams is produced by an internal version of Claude Tag. That is a production deployment, not a demo.

Where GitHub Copilot and Cursor are contained within the IDE, Claude Tag embeds in the communication hub — making AI task delegation accessible to non-technical team members as well as engineers. Available to Claude Enterprise and Team plan customers, with admin controls over token usage and audit logs of work history.

Source: ITmedia AI+ - ClaudeをSlackチャンネルに召喚、“チームの一員”として直接指示　新機能「Claude Tag」登場

Devin’s Japan Launch Reveals What Makes a “Real” Autonomous Agent — MicroVMs and a 1,582% Growth Claim

Cognition AI, maker of the AI coding agent Devin, held a business briefing shortly after establishing its Japan subsidiary in April 2026. The claimed 1,582% year-over-year user growth is eye-catching but meaningless without the absolute baseline. The concrete case studies are more useful.

DeNA deployed Devin across roughly 3,000 employees company-wide and completed a legacy system modernization project in approximately one month — a project originally scoped for six months. The compression factor is attributed largely to parallel task execution. Devin’s key technical differentiator is its “MicroVM” architecture: each AI agent runs in an isolated virtual machine and completes its sub-task self-sufficiently, allowing multiple concurrent sub-tasks to run without interference. This is the claimed separation from competitors.

Cognition was founded in 2023 and acquired AI coding editor Windsurf in 2025 to broaden its technical portfolio. Japan is its first Asian market, signaling high strategic priority. The enterprise focus is on legacy modernization and SI (systems integrator) productivity — with initial traction in finance and public sector.

The competitive field is crowded: Cursor, GitHub Copilot, Anthropic’s Claude Code, and others. Devin’s differentiation claim is end-to-end ownership from design through implementation and operations, positioning itself as an operating system for the entire engineering organization rather than a point solution. For startups, the remaining entry points appear to be SI-facing customization and vertical-specific variants.

Source: ITmedia AI+ - 国内ユーザー数「前年比1582％増」――AI開発支援「Devin」は競合と何が違う？

PixVerse and the Democratization of AI Video — What 100M Users and V6 Actually Deliver

Singapore-based PixVerse, founded in 2023, closed a Series C and joined the unicorn club in March 2026. At 100 million-plus users across 177 countries, it is among the largest AI video generation platforms by reach.

The standout evidence is in the user stories. A homemaker in Spain with no filmmaking background produced over 100 animated short films on PixVerse, grew a YouTube channel to 1 million subscribers and 10 million cumulative views, and has a full-length animated feature film — Aisha: Legend of the Desert — scheduled for release later in 2026. A Vietnamese TV network replaced a dangerous underwater taxi scene with AI-generated footage. These are production deployments, not proofs of concept.

The latest model, V6, generates video from text prompts with synchronized audio, multi-scene composition, up to 15 seconds at 1080p, with on-screen text support for Japanese, English, and Chinese. A “near real-time” generation mode is also claimed.

Competitors include Google Veo, ByteDance (Jianying/CapCut), and Runway, which entered the Japanese market in May. PixVerse’s positioning is speed and price. API integration is expanding, accelerating a B2B push. The broader market gravity is moving from standalone software sales toward platform and API models — a current PixVerse is well-positioned to ride.

Source: ITmedia AI+ - 主婦がAIアニメでYouTube登録者100万人を突破――1億ユーザーのAI動画生成サービス「PixVerse」の実態

Mercari Joins ChatGPT’s “Apps” — MCP Ecosystem Crosses Into Production

Mercari launched its official app in “Apps in ChatGPT” on June 23, built on the “Mercari MCP” (Model Context Protocol) infrastructure it open-sourced in January. The integration brings product search and listing assistance directly into ChatGPT’s conversational interface.

Two capabilities are in scope. On the search side, the integration handles open-ended natural language queries — “find camping gear under ¥5,000” or “birthday gift ideas for a soccer-loving kindergartner” — with multilingual search support. On the listing side, users describe an item and the system auto-generates a title, category, description, and suggested price benchmarked against comparable listings, with bulk draft creation for multiple items.

The more consequential layer is MCP itself. A shared protocol through which ChatGPT, Claude, and internal enterprise tools can connect to external services is maturing into a real standard. Mercari publishing Mercari MCP in January and shipping a ChatGPT integration within six months signals that the MCP ecosystem has moved beyond the experimental phase.

The business implication cuts both ways. Reaching users who never opened the Mercari app directly is the upside. The structural risk is that as more services connect via MCP, ChatGPT (or Claude) becomes the default UI — and platform-native interfaces erode. Every MCP integration is a bet that distribution through AI interfaces outweighs the value of owning the user session.

Source: ITmedia AI+ - 「Apps in ChatGPT」にメルカリ登場　自社MCP基盤を活用、会話で商品検索など

Takeaway

Today’s news draws a consistent throughline: AI profits are concentrating in vertically integrated players. OpenAI’s Jalapeño pushes toward self-sufficiency across model, service, and chip. Kioxia is harvesting AI demand at the infrastructure layer at a scale few anticipated. At the same time, Devin and PixVerse show that players who go deep on a specific use case and build compounding platform advantages can carve out durable market positions. The tension between full-stack integration pressure and focused vertical survival strategy is where the next generation of startup opportunity is likely hiding.