AI Just Broke the Semiconductor Industry's Forecast Model

Omdia just revised its 2026 semiconductor revenue forecast upward by 94.1% year-over-year. That is not a typo. The research firm, which tracks the global chip market, is now projecting that the semiconductor industry will nearly double its revenue this year, driven almost entirely by AI demand that has outstripped the industry’s physical capacity to build what the market wants.

The numbers are staggering on their own. Memory ICs alone are expected to account for more than 50% of total semiconductor revenue in 2026. Computing and data storage, the application market most directly tied to AI infrastructure, will rise by more than 150% year-over-year and approach $1 trillion in revenue. The semiconductor industry has had boom cycles before, but none driven by a single application category with this kind of force.

The Bottleneck Is Real and It Will Last

The most important detail in Omdia’s forecast is not the growth number. It is the constraint. AI demand has exceeded the industry’s current ability to produce and package chips. The bottlenecks are not theoretical. They exist across high bandwidth memory, advanced packaging, and node capacity, and Omdia expects them to persist until at least 2027. To put this in perspective, the semiconductor industry typically operates on 18-to-24-month capacity planning cycles. New fabrication plants take three to five years to build and equip. The demand surge that started in 2023 with the launch of ChatGPT and accelerated through 2024 and 2025 caught the industry in a position where expansion plans made before the AI boom were suddenly inadequate. Every major chipmaker is now scrambling to add capacity, but the physics of semiconductor manufacturing means that supply cannot respond as quickly as demand shifts.

This matters because it changes the economics of the entire AI supply chain. When demand outpaces supply for an extended period, prices rise, lead times stretch, and the companies with the deepest pockets and the strongest supplier relationships win. NVIDIA has been navigating this dynamic for two years, securing priority allocation from TSMC and SK Hynix. Smaller companies and newer entrants face a different reality, where the chips they need are either unavailable or available only at prices that compress their margins.

The bottleneck is particularly acute in high bandwidth memory, the specialized DRAM that AI accelerators like NVIDIA’s H100 and H200 require. HBM production requires advanced packaging techniques that are significantly more complex than standard DRAM manufacturing. SK Hynix and Samsung dominate HBM production, and both are running at full capacity. Micron is ramping up, but it will take time to reach the scale needed to ease the shortage.

Advanced packaging is the other chokepoint. TSMC’s CoWoS packaging technology, which stacks multiple chiplets on a single interposer, is essential for producing the largest AI accelerators. TSMC has been expanding CoWoS capacity aggressively, but demand keeps growing faster than supply. The result is a queue: companies that need the most advanced packaging are waiting months for allocation.

Memory ICs Take Over the Revenue Chart

The fact that memory ICs will account for more than 50% of total semiconductor revenue in 2026 is a structural shift, not a cyclical one. In previous chip booms, logic chips, processors, and analog components drove the majority of revenue. Memory was a commodity market, subject to wild price swings but never the dominant revenue category.

AI changed that. Training and running large language models requires enormous amounts of memory, both in the form of HBM for accelerator chips and standard DRAM for the servers that host them. Every new AI model is larger than the last, and every deployment requires more memory to serve it. The demand curve for memory is steeper than for any other chip category, and the supply response has not kept pace.

This has implications beyond the semiconductor industry. Higher memory costs flow directly into the price of AI servers, which flow into the cost of running AI workloads, which flow into the pricing of AI services. When memory prices rise, every layer of the AI stack gets more expensive. Cloud providers absorb some of that cost, but not all of it. The end result is that AI infrastructure costs more to build and operate than most organizations anticipated when they started their AI journeys.

The $1 Trillion Data Center

Computing and data storage approaching $1 trillion in revenue is a milestone that seemed distant even two years ago. The growth is being driven by data center servers, which are the physical infrastructure that runs AI models, and by memory-intensive applications that require high-density storage. The trillion-dollar figure is worth pausing on because it represents a fundamental rebalancing of where computing value resides. For decades, the semiconductor industry’s revenue was distributed across consumer electronics, automotive, industrial, and telecom applications. No single category dominated. AI has changed that math entirely. Computing and data storage now represents the largest single application market, and its share is growing, not stabilizing. This concentration creates risks. When one application category drives the majority of industry growth, a slowdown in that category has outsized effects on the entire supply chain. If AI investment cools, or if a breakthrough in model efficiency reduces the hardware required per unit of intelligence, the semiconductor industry could face a correction that is sharper than anything in its recent history. The industry has seen this pattern before, most memorably in the late 1990s when telecom infrastructure spending collapsed and took a significant chunk of chip demand with it. The difference this time is that AI demand is not just growing. It is compounding. Each generation of AI models requires more compute than the last. Each new application, from autonomous agents to real-time video generation to scientific simulation, adds demand on top of existing workloads. The compounding effect makes it harder for supply to catch up, even as chipmakers expand capacity at unprecedented rates.

The demand is not coming from traditional enterprise computing alone. AI training clusters, which can consist of thousands of GPUs connected by high-speed interconnects, represent a new category of compute demand that did not exist at scale before 2022. These clusters consume power, memory, and cooling at rates that dwarf traditional server deployments, and they need to be replaced or upgraded every two to three years as model sizes grow.

Cloud hyperscalers are the primary buyers. Microsoft, Google, Amazon, and Meta are collectively spending hundreds of billions of dollars on AI infrastructure in 2026, and much of that spending flows directly into semiconductor revenue. The scale of these investments has created a feedback loop: more AI infrastructure enables more AI applications, which generates more demand for AI infrastructure, which requires more chips.

What This Means for the AI Industry

The semiconductor constraint is not just a supply chain problem. It is a strategic variable that shapes which AI companies succeed and which ones struggle. Companies with guaranteed chip supply can train larger models, deploy them faster, and scale more aggressively. Companies without that access are forced to optimize around scarcity, using smaller models, more efficient architectures, or cloud-based inference instead of on-premises training. The uneven distribution of chip access is creating a two-tier AI industry. On one side, well-capitalized companies like OpenAI, Google, Anthropic, and Meta can secure the silicon they need to push the frontier. On the other side, startups and mid-size companies must compete for whatever capacity remains, often at premium prices that make their unit economics unworkable. This is not a temporary market imbalance. It is a structural feature of an industry where the most critical input has a limited supply and long lead times. The geopolitical dimension adds another layer of complexity. U.S. export controls have restricted which chips can be sold to Chinese companies, pushing China to develop its own semiconductor capabilities. Huawei’s Ascend line, manufactured by SMIC using older process nodes, represents an attempt to build an AI chip ecosystem outside the Western supply chain. It is less capable than NVIDIA’s latest hardware, but it exists, and it is improving. The semiconductor supply chain is no longer just a commercial question. It is a national security question, and the answers are being shaped in Washington and Beijing as much as in Silicon Valley and Hsinchu.

This dynamic is accelerating a consolidation trend in AI. The companies that can afford to build and operate large AI infrastructure are pulling away from those that cannot. The gap between the haves and have-nots in AI is, in many ways, a gap in semiconductor access.

The smartphone market offers a preview of what happens when memory costs rise across the board. Omdia notes that smartphone prices have been increasing since Q4 2025, particularly in the premium tier, where healthier margins can absorb inflated memory costs. This is narrowing the price gap between mid-tier and premium devices, pushing consumers toward higher-end phones. The same dynamic will play out in AI: organizations will either pay more for better infrastructure or settle for less capable systems.

Looking Ahead

From mid-2026 through early 2027, Omdia expects the semiconductor market to be defined by relentless AI demand. Capacity will remain constrained, advanced nodes will be heavily utilized, and memory and advanced packaging costs will continue to rise. Investment tied directly to AI infrastructure will drive record silicon consumption, while non-AI markets continue to face supply constraints.

The bottleneck will eventually ease. TSMC is expanding capacity. SK Hynix and Samsung are building new HBM lines. Micron is entering the market. But “eventually” is doing a lot of work in that sentence. For organizations building AI infrastructure today, the constraint is immediate, the costs are real, and the wait for relief is measured in years, not months.

The semiconductor industry is not just riding the AI wave. It is being reshaped by it. Memory companies are becoming more important than processor companies. Packaging technology matters as much as transistor design. Supply chain relationships determine competitive advantage more than engineering talent. The rules that governed the chip industry for decades are being rewritten in real time, and the new rules are still taking shape. The companies that navigate this reshape successfully will define the next decade of computing. The ones that do not will find themselves on the wrong side of the biggest infrastructure shift since the internet. The 94.1% growth number is not the story. The story is what happens when an entire industry discovers that its most important customer is growing faster than its ability to deliver.