AI Bottlenecks in 2026: Chips, HBM Memory, Packaging, and Power

The easiest argument to make would be: “Small models are eroding demand for data centers.” The numbers tell the exact opposite story. Tech giants’ capital expenditures are accelerating, not slowing down—but the real constraint is no longer silicon itself. It’s everything else needed to turn it into a functioning chip and keep it running.

AI investments are accelerating: Big Tech’s capex continues to grow

The combined capital expenditures of Microsoft, Amazon, Alphabet, and Meta for 2026 are nearing $725 billion, compared to about $410 billion in 2025—a 77% increase in just one year. Amazon leads with about $200 billion, followed by Microsoft ($190 billion), Alphabet (up to $205 billion, raised during the year), and Meta (up to $145 billion, with guidance raised twice).

Free cash flow is being hit hard—some estimates point to a decline of nearly 90% for Meta, with projections of negative free cash flow for Amazon in 2027–28. And this is where the first interesting signal for investors emerges: following Alphabet’s quarterly earnings report, the stock fell 7% despite 63% growth in its cloud business. Investors are beginning to demand visible ROI, not just blind growth in capex. It’s the first crack in a narrative that until now seemed beyond doubt.

Nvidia and AMD: Demand for GPUs and AI chips isn’t slowing down

If customer capex accelerates, it’s first evident in the revenue of the companies selling them GPUs. Nvidia closed the quarter with $75.2 billion in data center revenue, up 92% year-over-year, with a net margin of 72%. AMD follows with $6.7 billion in the data center segment, up 107% year-over-year, and total revenue up 50%, driven by the ramp-up of Helios: AMD’s “rack-scale” system—an entire server cabinet designed to have hundreds of chips work together as a single large computer for AI, in direct competition with Nvidia’s equivalent solutions. Close ✕ (its rack-scale system) with Meta, OpenAI, and Oracle as customers.

Neither company shows any signs of slowing down due to the adoption of SLMs or local inference. If anything, the issue they discuss in calls with analysts is the opposite: they can’t deliver fast enough. And this is where the narrative shifts from “how many chips are needed” to “what’s really needed to turn a chip into a working product.”

The three bottlenecks in the AI supply chain: packaging, memory, and power

The public narrative almost always focuses on the shortage of GPUs. But by 2026, the constraint has shifted downstream to three specific points in the chain—and it is these, not the adoption of smaller models, that are the true limit to how fast AI infrastructure can grow.

1. Advanced packaging and CoWoS: why assembling AI chips has become a bottleneck

A modern AI chip is not a single piece of silicon: it is a logic processor bonded to stacks of ultra-high-speed memory through a process called CoWoS (Chip-on-Wafer-on-Substrate), controlled almost entirely by TSMC. Without this step, even a perfect 3-nanometer wafer remains a piece of inert silicon, not a functioning chip.

Nearly four times the capacity in two years — and yet TSMC CEO C.C. Wei told shareholders that CoWoS capacity remains “extremely tight and sold out through all of 2026.” Nvidia alone has reserved an estimated 800,000 to 850,000 wafers of 2026 capacity—more than half of the total available. According to some industry estimates, TSMC’s advanced packaging facilities are booked through 2027, with lead times ranging from 52 to 78 weeks. ABF substrates—another essential material, over 95% of which is controlled by a single Japanese supplier—are also cited by TSMC as the next potential bottleneck after memory. It’s worth noting this, as it may represent the tightest point in the entire supply chain. ABFs (Ajinomoto Build-up Film) are the essential insulating material used to encapsulate every high-end AI chip—and the name is no coincidence: they were discovered in the 1990s as a byproduct of monosodium glutamate production, the common bouillon cube. Ajinomoto, the Japanese company best known to the general public for its food products, now controls over 95% of the global market for this material—a concentration even more extreme than TSMC’s in advanced packaging. In August 2026, it notified its Chinese customers of a 30% cut in supplies and a new pricing structure effective in the third quarter, with increases of up to 30%; delivery times exceeded six months. According to Morgan Stanley, the gap between supply and demand for high-end ABF substrates will rise from 10% in the second half of 2026 to 21% in 2027, reaching 42% in 2028. The company has announced investments of at least 25 billion yen by 2030 to increase capacity by 50%—but with Goldman Sachs forecasting compound annual market growth of 33% through 2028, it is by no means certain that this will be enough. The stock, historically viewed as a low-profile food company, has gained over 40% since the start of 2026, hitting an all-time high in late February: the market has begun—albeit a few months late—to realize that a condiment company holds the entire AI chip supply chain by the throat.

2. HBM and DRAM Memory: Why AI Is Absorbing Global Production Capacity

Here, the bottleneck has become visible even to those who don’t follow the stock markets. SK Hynix, Samsung, and Micron—which together control over 90% of global DRAM production—have announced that their HBM (High Bandwidth Memory) capacity is sold out for all of 2026, with some customers having already secured supplies through 2027. The reason: producing a single HBM wafer requires about three times the capacity of a conventional DRAM wafer, and manufacturers have shifted production en masse toward HBM, which is much more profitable.

Samsung has raised the price of 32GB DDR5 modules from $149 to $239—a 60% increase—while contract prices for DDR5 have more than doubled over the same period. This isn’t a problem for industry insiders: it’s already factored into the price of the next computer you’ll buy.

Forecasts do not point to a quick rebound: Samsung and SK Hynix have indicated that the shortage could last at least until 2027, and some statements from SK Group have even mentioned sustained demand pressure through 2030. By 2026, data centers are expected to account for up to 70% of all high-end memory produced worldwide—a near-complete reversal for an industry historically geared toward consumer PCs and smartphones.

3. Energy and the Power Grid: The Physical Limit to AI Data Center Growth

The final bottleneck is also the most difficult to resolve through investment in a new factory: the power grid. A single AI training campus can require between 100 and 500 megawatts of continuous power—equivalent to the needs of a small city. Wait times for connecting new capacity to the grid have lengthened to 3–5 years, compared to the 24–30 months typical before 2020, primarily due to a shortage of high-voltage transformers.

The result: according to an analysis cited by several industry players, of the approximately 12 gigawatts of data center capacity announced for 2026 in the United States, only 5 are actually under construction. The rest remain stuck in the planning phase. The five largest Big Tech companies are now collectively spending—on energy capital expenditures alone—nearly double what the entire U.S. electric power industry invested in generation, transmission, and distribution in a recent year. It is not surprising that in some residential areas near large clusters of data centers (Virginia, Texas, Georgia), electricity bills have already risen by 8–15%, nor that analysts like Gartner predict that by 2027, power constraints will limit up to 40% of planned AI data centers.

Big Tech’s response has been to bypass the public grid entirely: direct agreements with energy producers (Microsoft signed a deal for 10.5 GW with Brookfield Renewable), investments in modular nuclear power plants, and on-site fuel cells. This is a sign worth reading carefully: when the world’s most highly capitalized companies start building their own power plants rather than wait for the grid, the bottleneck is no longer perceived as temporary.

Emerging AI Chips: Microchip, CEVA, and Cerebras Between the Edge and the Cloud

Amid this race dominated by a few mega-caps, there are smaller stories worth watching, because they reveal the other side of the same coin: not everyone is betting on the same scale.

Microchip Technology acquired Hailo, an Israeli startup specializing in edge AI chips, after the latter saw its valuation plummet from over a billion dollars to a fraction of that amount—not a triumphant acquisition, but a rescue, integrated into the OEM distribution network that Microchip has built over decades. CEVA, an IP licensor for NPUs on a much smaller scale, has just signed a licensing agreement with what it calls “a leading global AI platform” — a design win that, if it translates into royalties from hyperscaler volumes, could transform the company’s scale within 2–3 years. And Cerebras, with its $6.4 billion IPO, is more accurately positioned as a challenger to Nvidia in ultra-high-speed cloud inference—not as proof of the “edge” thesis: its cloud revenue grew 281% year-over-year, but its GAAP net loss in the last quarter alone exceeded $450 million.

The future of AI will depend on packaging, memory, and power

When you put together the financial statements of Big Tech, chipmakers, and outsiders, the picture that emerges isn’t the one suggested by the easiest headline to write. There’s no evidence that the rise of SLMs is cannibalizing demand for data centers—on the contrary, capex and revenue are accelerating on both fronts simultaneously. The real investment thesis, for those who follow this sector, isn’t “edge versus cloud”: it is that the bottleneck has shifted from the chip to everything needed to make it work—packaging, memory, and electricity—and that these three physical constraints, not the adoption of smaller models, are what will truly determine the speed at which AI infrastructure can grow in the coming years.

It’s the same underlying principle as in the previous article, viewed from a different angle: artificial intelligence is diverging between an ever-larger, more powerful cloud and an ever-more-capable edge—not one at the expense of the other, but both constrained by the same fundamental physics: how much memory can be produced, how many chips can actually be assembled, and how much electricity can be delivered where it’s needed.