Key Takeaway: July 2026 marks a structural shift in the AI industry. Four of the world’s most prominent AI companies — OpenAI, DeepSeek, Anthropic, and Meta — are all building custom inference chips to reduce dependency on Nvidia and optimize for their specific AI architectures. Broadcom has emerged as the design partner of choice, while TSMC remains the sole foundry for all custom AI silicon. The era of “buying compute” is giving way to “defining compute.”
For years, the dominant narrative in artificial intelligence was straightforward: train your models on Nvidia GPUs, pay per token for inference, and let the world’s most valuable semiconductor company handle the silicon. That model is now breaking apart. In the span of just a few weeks in June and July 2026, four AI giants have advanced their custom chip programs — each driven by the same economic reality: at scale, general-purpose GPUs are becoming too expensive and too inefficient for model-specific inference workloads.
This article examines what each company is building, why it matters, and what it means for the broader technology ecosystem — including the embedded systems and electronics industries that power the infrastructure behind AI.
OpenAI Jalapeño: The Fastest ASIC Development Cycle in History
On June 24, 2026, OpenAI unveiled Jalapeño, its first custom Intelligence Processor — an accelerator designed from the ground up for large language model inference. The chip was co-developed with Broadcom and Celestica in just nine months, a timeline OpenAI describes as potentially the fastest ASIC development cycle ever achieved in high-performance semiconductors.
What makes Jalapeño significant isn’t just the speed of development — it’s the design philosophy. Unlike Nvidia’s general-purpose GPUs, which must support training, inference, and a wide range of parallel computing tasks, Jalapeño is optimized specifically for the transformer architecture that underpins GPT and its successors. Every transistor, every memory interface, every compute lane is tuned for the specific patterns of attention mechanisms and token generation.
For embedded systems engineers, Jalapeño represents a broader trend: application-specific integrated circuits (ASICs) are becoming viable at timelines and cost points that were unthinkable just two years ago. Broadcom’s ability to take a chip from concept to tape-out in nine months suggests that custom silicon is no longer the exclusive domain of companies with billion-dollar R&D budgets.
DeepSeek: Building Silicon Under Sanctions
On July 7, 2026, Reuters reported that Chinese AI company DeepSeek is developing its own AI inference chip, with the project having started approximately one year ago. The chip is designed specifically for inference — the process of generating responses to user queries — not for training.
DeepSeek’s journey to custom silicon is a story of adaptation under constraint. The company trained its R1 model on Nvidia H800 chips before they were banned from export to China. It then adapted its V4 model to run on Huawei’s Ascend architecture. Now, it’s designing its own silicon entirely.
The strategic implications are significant. With an estimated $7.4 billion in funding backing the effort, DeepSeek’s custom inference chip represents China’s most serious attempt to build competitive AI silicon independent of both Western suppliers and Huawei’s ecosystem. If successful, it would demonstrate that AI chip design expertise is diffusing globally, not concentrating in Silicon Valley.
For the electronics manufacturing ecosystem in India and Southeast Asia, DeepSeek’s chip development raises important questions about supply chain diversification and the future geography of AI hardware production.
Anthropic Explores Samsung for Custom Silicon
On July 2, 2026, The Information reported that Anthropic is in early discussions with Samsung Electronics about potential manufacturing partnerships for custom AI chips. The project remains in very early stages — Anthropic has not yet finalized the chip’s design, processing function, performance specifications, or deployment architecture.
According to sources, Anthropic is considering Samsung’s 2nm process technology and advanced packaging solutions. The company has engaged with multiple chip design firms but has not yet entered detailed design, testing, or manufacturing phases.
What’s notable about Anthropic’s approach is its multi-supplier hedging strategy. While OpenAI is increasingly tied to Nvidia and xAI operates its own infrastructure, Anthropic has been building relationships across AWS Trainium, Google TPU, and Nvidia GPUs. Adding custom silicon from Samsung would be a fourth option — a deliberate effort to avoid lock-in to any single hardware provider.
Anthropic’s official response emphasizes that AWS Trainium, Google TPU, and Nvidia GPUs will remain central to its compute strategy. The custom chip effort is about building optionality, not replacing existing infrastructure.
Meta Iris: Production-Ready in September 2026
On July 9, 2026, Reuters reported that Meta’s custom AI chip, code-named Iris, will enter production in September 2026. An internal memo reviewed by Reuters shows the chip cleared six weeks of testing without any major issues — a notable milestone for a program that has struggled for years.
Iris is part of Meta’s Training and Inference Accelerator (MTIA) program, which includes four planned chip generations: MTIA 300, 400, 450, and 500. The company plans to release a new generation approximately every six months through 2027, roughly twice the speed of a typical chip development cycle.
Broadcom is the design partner on Iris, and TSMC is the foundry. Meta has also signed long-term supply agreements with Samsung Electronics for memory chips, Sandisk for flash storage, and Sumitomo Electric for fiber-optic equipment.
Meta’s internal memo contains a candid admission: adopting the latest GPUs at Meta’s scale “has been a heavy lift, and it has cost us time.” Iris is designed to run ranking and recommendation systems (which power Facebook and Instagram feeds) and generative AI features — alongside, not instead of, the GPUs Meta continues to buy from Nvidia and AMD.
The infrastructure push behind Iris is staggering. Meta plans to double its data center computing capacity from 7 gigawatts in 2026 to 14 gigawatts in 2027. To put that in perspective, 14 gigawatts would power more than 11 million homes.
Broadcom: The Quiet Winner of the Custom Chip Race
While headlines focus on the AI companies building chips, the common thread across three of the four major custom silicon programs is Broadcom. The chip design firm is the partner behind Google’s TPU, OpenAI’s Jalapeño, and Meta’s Iris.
The financial numbers tell the story:
- Q2 2026 AI semiconductor revenue: $10.8 billion, up 143% year-over-year
- Q3 2026 AI revenue guidance: $16.0 billion, over 200% growth
- Fiscal 2027 AI revenue forecast: Expected to exceed $100 billion
Broadcom’s role as the go-to custom chip design partner is built on a specific value proposition: it can take a company’s AI workload requirements and translate them into optimized silicon without requiring the customer to build a chip design team from scratch. This is the fabless model applied to AI accelerators, and it’s proving to be enormously profitable.
For the broader electronics industry, Broadcom’s success validates a design-services-plus-foundry model that could eventually extend beyond AI into industrial automation, automotive, and telecommunications — any domain where application-specific silicon provides a competitive advantage over general-purpose processors.
The TSMC Factor: Everyone’s Chips Come From One Foundry
Perhaps the most striking aspect of the AI custom chip race is that all four companies’ chips — regardless of who designs them — are manufactured by TSMC. This means the AI chip design battle is fought on software and architecture, not manufacturing.
TSMC’s dominance is reflected in its financial results. June 2026 revenue surged 67.9% year-over-year to NT$442.68 billion, with first-half 2026 cumulative revenue up 35.6%. The company’s Q2 earnings report, due July 16, is expected to confirm that AI-driven demand continues to accelerate.
The concentration of all custom AI silicon at a single foundry creates both opportunity and risk. For chip designers, TSMC’s manufacturing excellence means they can compete on architecture rather than process technology. For the global supply chain, it means a single point of failure — a reality that governments and corporations are increasingly uncomfortable with.
Memory Chips: The “Golden Child” of the AI Revolution
The AI custom chip race is unfolding against the backdrop of a severe memory chip shortage. Wedbush Securities tech strategist Dan Ives argued on CNBC that memory chips have become the most valuable slice of the AI supply chain, calling them the “golden child” of the AI revolution.
Ives quantified the imbalance: demand outstrips supply by approximately 15 to 1, a gap he expects to persist until at least 2028. The numbers are staggering:
- Micron Technology: Stock up 696% over the past year, market cap $1.1 trillion. Q3 revenue reached $41.5 billion, up 345% year-over-year, with gross margin expanding to 84.6%.
- SK Hynix: Completed its U.S. listing debut on July 11, 2026 — a signal that capital is rotating toward American-listed AI infrastructure names.
- High-Bandwidth Memory (HBM): The choke point in the AI buildout, as Nvidia’s Blackwell accelerators rely on stacked DRAM modules that only SK Hynix and Samsung currently produce at scale.
For electronics manufacturers in India and elsewhere, the memory shortage has direct implications for component procurement, pricing, and product planning. Any device that uses DRAM or NAND — from industrial controllers to consumer electronics — is facing upward price pressure.
What This Means for Embedded Systems and Electronics Engineers
The AI custom chip race isn’t just a story for Wall Street analysts and semiconductor investors. It has practical implications for embedded systems engineers and electronics manufacturers:
- ASIC development timelines are shrinking. OpenAI’s nine-month Jalapeño development cycle suggests that custom silicon is becoming more accessible. This could eventually trickle down to industrial and automotive applications.
- Broadcom’s design-services model may expand. If application-specific silicon proves viable for AI, the same approach could be applied to motor control, sensor processing, and industrial automation — domains where custom logic outperforms general-purpose MCUs.
- Memory pricing will remain elevated. The 15:1 demand-supply imbalance means component costs will stay high through 2028, affecting BOM budgets across the electronics industry.
- Geopolitical risk is real. DeepSeek’s custom chip development under sanctions demonstrates that AI hardware is becoming a strategic asset, with implications for export controls and supply chain planning.
Related Reading
- TI’s Single-Chip Multi-Axis Motor Control — 500ns Latency for Robotics
- Technologies — More industry analysis on justLast
Frequently Asked Questions
Why are AI companies building their own chips instead of using Nvidia GPUs?
The shift is driven by inference economics. As AI moves from training to inference at scale, general-purpose GPUs become too expensive and inefficient for model-specific workloads. Custom ASICs deliver better performance per watt and lower cost per inference token at the scale these companies operate.
Who manufactures all the custom AI chips?
All four major custom AI chips — OpenAI’s Jalapeño, Meta’s Iris, Anthropic’s (future) chip, and DeepSeek’s (future) chip — are or will be manufactured by TSMC (Taiwan Semiconductor Manufacturing Company). This means the competitive advantage lies in chip architecture and software, not manufacturing capability.
What is Broadcom’s role in the AI custom chip race?
Broadcom is the design partner for three of the four major custom AI chip programs: Google’s TPU, OpenAI’s Jalapeño, and Meta’s Iris. The company provides end-to-end chip design services, from architecture to tape-out, allowing AI companies to build custom silicon without establishing their own chip design teams.
How long does it take to develop a custom AI chip?
OpenAI’s Jalapeño was developed in approximately nine months — potentially the fastest ASIC development cycle in history. Meta’s Iris has been in development for several years as part of the MTIA program. DeepSeek’s chip has been in development for about one year. The timeline is compressing as design tools and foundry processes mature.
What is the AI memory chip shortage?
The demand for high-bandwidth memory (HBM) used in AI accelerators far exceeds supply — analyst Dan Ives estimates a 15-to-1 demand-supply imbalance. Micron, SK Hynix, and Samsung are the primary HBM suppliers, with production capacity constrained through at least 2028.
Does this affect the Indian electronics manufacturing industry?
Yes, in several ways. The memory shortage affects component pricing for all electronics manufacturers. The ASIC development trend could eventually create opportunities for Indian semiconductor design services. And the geopolitical implications of custom chip development (particularly DeepSeek’s effort under sanctions) affect global supply chain planning.
This article was originally published on justLast.in on July 14, 2026. Sources: Electrek, Tom’s Hardware, Reuters, The Information, 247WallSt, CNBC, gzmato.com, SemiEngineering.
