Semiconductor Watch: The AI Chip Race, NVIDIA: The Unassailable Leader, AMD: The Credible Challenger, & Intel: The Foundry Revival
The CODEW Semiconductor Watch | August 26, 2026
The semiconductor industry has entered a new era. In 2026, the global semiconductor market is projected to surpass $1.5 trillion — a staggering 90% year-over-year growth driven entirely by AI infrastructure demand. The AI chip market is on a trajectory toward $1 trillion, with NVIDIA still dominant but facing challenges from every direction: AMD's MI400 series, Intel's foundry revival, hyperscaler custom silicon, and a wave of specialized startups. This Watch examines the shifting competitive landscape, the bottlenecks that still constrain supply, and what comes next for the semiconductor industry's most consequential cycle.
Executive Brief
The semiconductor industry has transitioned from the speculative "AI gold rush" into a mature, trillion-dollar infrastructure era. The sector pulse is defined by a widening divergence between NVIDIA and its challengers, but the emergence of specialized silicon and the successful ramp of Intel's 18A node have introduced new variables into a previously one-sided equation.
By late 2025, NVIDIA's Blackwell (B200/GB200) platform accounted for over 80% of its high-end shipments, driving data center revenue to a record $51.2 billion in a single quarter. Analysts estimate NVIDIA's Q2 2026 revenue will nearly double year-over-year to approximately $92 billion — more than most competitors' annual revenue.
But the competitive landscape is shifting. JPMorgan forecasts that AI ASIC/XPU shipments could exceed GPU shipments by 2027 — approximately 12.5 million ASICs versus 10.9 million GPUs. Custom silicon shipments are growing at 44.6% year-over-year, nearly three times the 16.1% growth rate of GPUs. The era of a single dominant architecture is ending.
Semiconductor Market This Week
The semiconductor market reached historic milestones this week:
- Global semiconductor market: Projected to reach $1.5 trillion in 2026, four years ahead of prior forecasts
- AI chip market: On track toward $1 trillion, with AI chips accounting for 30% of all chip sales
- Top five chipmakers (NVIDIA, Intel, Broadcom, AMD, Qualcomm): Combined sales expected to exceed $538 billion in 2026
- Hyperscaler capex: AWS, Google, Microsoft, and Meta have collectively guided to more than $300 billion in 2026 capital spending, with AI infrastructure accounting for the majority
📊 THE CODEW STAT
$1.5 trillion — Global semiconductor market in 2026 (90% YoY growth)
$92 billion — NVIDIA's estimated Q2 2026 revenue
44.6% — Custom AI chip shipment growth (vs. 16.1% for GPUs)
The AI Chip Race
NVIDIA: The Unassailable Leader
NVIDIA maintains approximately 80-85% market share in AI chips, with dominant positioning in both training and inference. The company's software ecosystem — CUDA — remains the industry's deepest moat, with developers trained on CUDA preferring to stay on NVIDIA hardware.
NVIDIA's product portfolio spans multiple generations. The H100 remains the industry workhorse with 80GB HBM3 memory and 3,350 GB/s bandwidth. The H200 extends memory to 141GB HBM3e with 4,800 GB/s bandwidth. The B200, NVIDIA's latest generation, offers 192GB HBM3e memory, 8,000 GB/s bandwidth, and 2x+ tensor performance over H100.
Vera Rubin — the successor to Blackwell — is already in production for a late 2026 debut. NVIDIA's data center revenue reached $75.2 billion last quarter, with gross margins remaining in the low-to-mid 70% range.
AMD: The Credible Challenger
AMD has successfully carved out a significant "second-source" position, securing a double-digit share of the AI accelerator market for the first time. The company's Instinct MI350 series, launched in mid-2025 on the 3nm CDNA 4 architecture, features up to 288GB of HBM3E memory and supports models up to 520B parameters on a single GPU. AMD's ability to offer a 30% discount relative to NVIDIA's top-tier chips while matching performance in specific LLM inference tasks has attracted Tier-1 customers like Meta and Microsoft.
The MI400 series and Helios AI rack, launching in 2026, will offer up to 10x more performance for frontier models, with 432GB HBM4 and 2.9 exaflops FP4 performance. The MI400 will deliver up to 40 petaflops FP4, 20 petaflops FP8, 432GB HBM4 memory, and 19.6 TB/s bandwidth. AMD's data center revenue reached $6.72 billion last quarter, and CEO Lisa Su projects the AI accelerator market will expand to $1.4 trillion by 2030.
Intel: The Foundry Revival
Intel has reached a critical milestone: high-volume manufacturing of the 18A (1.8nm) process node at the end of 2025. This represents the first time in over a decade that Intel has arguably reclaimed the transistor-density crown. The node utilizes RibbonFET (gate-all-around) transistors and PowerVia (backside power delivery) technology.
On the product side, Intel recognized it could not beat NVIDIA in raw AI training performance and pivoted its Gaudi roadmap to focus on enterprise AI inference. Gaudi 3 is capturing market share as a cost-efficient alternative for inference workloads. Intel's Crescent Island GPUs, focused on inference, are expected to ship this year. However, Intel's AI accelerators are still struggling to break the 10% market share barrier in a data center world standardized on CUDA.
Custom Silicon: The Hyperscaler Insurgency
The most significant structural shift in the AI chip market is the rise of custom silicon from hyperscalers. AWS Trainium 2, Google TPU v5p and Trillium, Microsoft Maia, Meta MTIA, and OpenAI's custom chip are reshaping where AI compute margin lives.
AWS has stood up Project Rainier for Anthropic at a scale of more than one million Trainium 2 chips. Google has placed TPU v5p into broad external availability and is shipping the inference-oriented Trillium (TPU v6e). Microsoft is ramping Maia 100 inside its own data centers for OpenAI and first-party workloads. OpenAI's Jalapeño chip, developed with Broadcom, is now outperforming NVIDIA's GB300 in internal tests — a significant milestone in the company's strategy to reduce dependence on NVIDIA.
JPMorgan forecasts that by 2027, AI ASIC shipments will surpass GPUs — 12.5 million versus 10.9 million. In 2026, ASICs are expected to account for approximately 42% of global AI chip shipments, with shipments growing 109% year-over-year, far exceeding GPU growth of 39%.
The ASIC advantage is most pronounced in inference: model parameters are fixed, low latency is critical, and ASICs offer superior cost-per-token economics. For training, GPUs retain an advantage due to their flexibility and mature software ecosystem.
Foundry & Manufacturing
TSMC: The Binding Constraint
TSMC remains the indispensable supplier of advanced AI chips. The company's CoWoS advanced packaging capacity is the industry's most critical bottleneck. TSMC's monthly CoWoS capacity is projected to increase from approximately 70,000 wafers in 2025 to 130,000–140,000 wafers by the end of 2026, and further to 220,000 wafers in 2027.
NVIDIA has reportedly reserved approximately 800,000 to 850,000 wafers of TSMC's CoWoS capacity for 2026, accounting for more than 50% of the company's total CoWoS capacity. Even with expansion, the CoWoS supply-demand gap will remain at approximately 20% in 2026.
To address capacity constraints, TSMC is expanding outsourcing of CoW (Chip-on-Wafer) — a key step in CoWoS packaging — to OSAT companies including ASE, Amkor, and SPIL. OSAT providers are placing equipment orders for new production lines.
ASML: The Enabler of the AI Era
ASML raised its full-year 2026 revenue forecast to €43-45 billion, its second increase this year, on strong AI chip demand. The company plans a 30% capacity expansion in each of the next two years for its flagship EUV tools. ASML's Q2 2026 report reinforces that demand for advanced chip manufacturing equipment remains very high.
HBM & Memory: The New Bottleneck
The defining bottleneck of the AI boom is no longer the GPU — it is the memory that sits next to it. "AI started as a compute race," writes 22V Research's Jordi Visser. "It is becoming a memory race".
The HBM market is highly concentrated. In Q1 2026, SK Hynix led with approximately 58% revenue share, followed by Samsung at 21% and Micron at 21%. The three makers together command more than 95% of the HBM market.
Micron reported fiscal Q3 2026 revenue of $41.46 billion, more than quadruple a year earlier, with management guiding Q4 to approximately $50 billion. CEO Sanjay Mehrotra said memory has become "a strategic asset" in the AI era. Micron has signed 16 strategic customer agreements representing roughly $100 billion in cumulative revenue through 2030.
SK Hynix began sampling HBM4 in early 2026. Samsung announced HBM4 mass production in February 2026. At Hot Chips 2026, Samsung proposed ZHBM — a concept that vertically stacks HBM directly on top of the XPU, targeting a 70% reduction in total DRAM power consumption compared to HBM5.
AI Networking & Data Center Chips
As AI clusters scale from thousands to hundreds of thousands of GPUs, the networking layer has become a critical bottleneck. Broadcom is the dominant player in AI networking, with its Tomahawk 6 switch chip now in mass production. The chip, built on TSMC 3nm, supports 64 800G ports or 512 200G ports and is optimized for both scale-out and scale-up AI networks.
Broadcom's AI networking revenues represented roughly one-third of total AI revenues in Q1 FY2026 and are expected to rise toward 40% as demand for higher-bandwidth infrastructure accelerates.
NVIDIA's Spectrum-X Ethernet platform can now connect up to 512,000 GPUs in a single system. The company's Groq 3 LPX low-latency inference racks, each integrating 256 Groq 3 chips and processing 3,400 tokens per second, entered full production.
AI Chip Startups & Emerging Challengers
The independent AI chip startup landscape is consolidating. Cerebras went public on May 14, 2026, closing day one at approximately a $56 billion valuation. The company's wafer-scale chip, the size of a dinner plate, claims to process AI prompts faster than NVIDIA.
Groq raised $350 million at a $3.5 billion valuation as it pivoted from chipmaker to Nvidia-powered neocloud. Tenstorrent has attracted acquisition interest from Intel and Qualcomm at a valuation potentially exceeding $5 billion. Cerebras and Tenstorrent remain the two primary standalone players in the space following NVIDIA's absorption of Groq's assets.
Jon Peddie Research estimates that approximately 150 companies are developing over 200 different AI semiconductor designs, targeting faster or more efficient inference chips. However, most of these companies are targeting inference — a more fragmented market than training.
Edge AI & Robotics Silicon
The AI chip market extends beyond data centers. Qualcomm introduced new edge AI processors, including the Dragonwing Q-8750 with 77 TOPS of AI performance. NXP introduced the i.MX 93W, the first chip to pair a dedicated AI neural processing unit with secure tri-radio wireless connectivity.
NVIDIA announced the Jetson Orin Nano 2, delivering 78 TOPS of AI performance with 8GB of memory — twice the inference performance of its predecessor. The module and developer kit are expected in the first half of 2027.
Supply Chain & Geopolitical Risk
The semiconductor supply chain remains fraught with geopolitical risk. U.S. export controls prohibit NVIDIA from selling its most advanced chips to China, accelerating Chinese development of domestic alternatives. U.S. senators have urged the administration to tighten regulations on TSMC and other foundries to prevent Chinese companies from obtaining advanced AI chips through overseas subsidiaries.
China's domestic AI chip ecosystem is advancing rapidly. Huawei's domestic AI chip share in China has risen to an estimated 50%. Meanwhile, Japanese semiconductor manufacturer Rapidus has started trial production for 2nm GAA transistors — a critical milestone in Japan's return to advanced logic chip manufacturing.
The CODEW Analysis
Is the AI semiconductor market moving beyond the GPU toward a broader ecosystem of custom silicon, advanced memory, networking, and specialized AI compute?
The answer is a definitive yes — and the transition is accelerating faster than most industry observers anticipated. The $1.5 trillion semiconductor market, the rise of ASICs to 42% of AI chip shipments, and the $100 billion in long-term HBM supply agreements all point to the same conclusion: the AI chip market is diversifying away from a single architecture.
Three trends are driving this shift. First, hyperscaler custom silicon is reaching scale. AWS's Project Rainier at one million Trainium 2 chips, Google's TPU v5p external availability, and Microsoft's Maia 100 ramp are not experiments — they are production deployments at meaningful scale. JPMorgan's forecast that ASICs will surpass GPUs by 2027 is increasingly credible.
Second, memory has become the binding constraint. The HBM market is a three-player oligopoly where SK Hynix, Samsung, and Micron command 95% of supply. Micron's $100 billion in long-term customer agreements suggests that enterprises view advanced memory as a strategic resource, not a commodity. The bottleneck has shifted from "who can build the best GPU" to "who can secure the memory and packaging to deploy it."
Third, the foundry and packaging ecosystem is under unprecedented strain. TSMC's CoWoS capacity is the industry's most critical bottleneck, with NVIDIA alone reserving over 50% of 2026 capacity. The CoWoS supply-demand gap will remain at 20% in 2026, and ASML's 30% capacity expansion reflects the scale of investment required to keep pace.
For enterprise AI buyers, the implication is clear: the era of treating AI compute as a single-architecture commodity is ending. The future belongs to multi-architecture strategies that balance GPUs, custom ASICs, and inference-optimized silicon across training and inference workloads. The companies that can navigate this complexity — securing memory supply, managing packaging constraints, and optimizing across architectures — will have a structural advantage in the next phase of AI deployment.
What to Watch Next
- NVIDIA Q2 earnings (Aug 26): The most anticipated event of the week. Watch for revenue, guidance, and commentary on the Blackwell transition and Vera Rubin timeline.
- TSMC CoWoS capacity expansion: Whether TSMC can reach 140,000 wafers per month by year-end and whether the supply-demand gap narrows.
- HBM4 production ramp: Whether SK Hynix, Samsung, and Micron can scale HBM4 production to meet surging demand.
- ASIC vs. GPU shipment data: JPMorgan's forecast that ASICs will surpass GPUs by 2027 — whether Q3 2026 data validates the trend.
- OpenAI Jalapeño chip deployment: Whether OpenAI's custom chip moves from internal testing to production use and how it affects NVIDIA's data center revenue.
- Export control enforcement: Whether the U.S. tightens regulations on foundries serving Chinese customers and how TSMC responds.
Source Attribution
- Wedbush — The Great Decoupling: How NVIDIA's AI Hegemony is Redefining the Semiconductor Hierarchy in 2026 (January 6, 2026)
- DeployBase — AI Chip Wars: NVIDIA vs AMD vs Custom Silicon 2026 Update (January 20, 2026)
- Quartr — AMD Advancing AI Keynote Summary (July 8, 2026)
- Wedbush — The 1.8nm Redemption: Intel Fights for Relevancy in a World Owned by Nvidia and Broadcom (April 1, 2026)
- Deluair — The Custom Silicon Insurgency Against Nvidia in 2026
- TrendForce — TSMC Reportedly Expands Outsourcing of Key CoWoS Front-End Step to OSATs (August 5, 2026)
- Longbridge — 'AI Equals Memory': Why The AI Boom Now Runs Through Micron, SK Hynix And Samsung
- Hashrate Index — Three Independent AI Chip Companies Taking On NVIDIA (May 15, 2026)
- Counterpoint Research — CES 2026 Edge AI Announcements (January 8, 2026)
- Economic Times — ASML capacity upgrade soothes AI chip bottleneck fears (July 16, 2026)
- JPMorgan — AI Accelerator Shipment Forecast 2026-2027
- WSTS / SIA — Global Semiconductor Market Projections 2026
Reviewed by Erwin Castro
on
Wednesday, August 26, 2026
Rating: