Beneath the flood of capital and public demonstrations in AI lies a very concrete industrial reality: production lines, specialised components, memory supply, packaging processes and power availability. While high-profile figures such as Jensen Huang, Sam Altman and Tim Cook attract attention with bold statements, the actual pace of progress is set by physical infrastructure.
How much is being invested — and what money cannot immediately fix
Big U.S. tech companies’ AI infrastructure spending reached roughly $630–650 billion in 2024. That scale of investment is significant but cannot instantly eliminate manufacturing and logistical bottlenecks: deploying new lithography systems cannot be rushed overnight, HBM memory capacity does not appear instantly, and energy or network expansions are not shortened by a single decision.
Key bottlenecks
-
Nvidia: IDC measured an 85.2 percent market share for Nvidia in AI accelerators in Q2 2025. This level of concentration indicates that a core area of AI hardware has coalesced around a single player, reinforced by Nvidia’s CUDA software ecosystem and the industry’s familiarity with it.
-
ASML: the Dutch company is the sole maker of EUV lithography systems. High-NA EUV units cost on the order of $400 million each, and ASML itself says broader industrial integration will take years — a constraint on rapid mass scaling.
-
TSMC: Taiwan Semiconductor Manufacturing Company plans $52–56 billion of investment for 2026 and anticipates roughly 30 percent growth. TrendForce reported that TSMC accounted for 69.9 percent of the global foundry market in 2025, underscoring how much of the semiconductor supply chain is concentrated among a few players.
Logistical and component-level frictions
Chips alone are not enough: it matters how they are packaged (advanced packaging, chip stacking), what memory is paired with them, whether optical data-transfer components are available, and whether sufficient power is on hand. A Broadcom executive said in late March that TSMC’s capacity had become a bottleneck for 2026; certain lasers and optical components now face waits measured in months rather than weeks. SK Hynix’s nearly $8 billion ASML order shows that memory makers are trying to secure capacity in advance.
Energy and networks: little discussed but highly consequential
Data centres run on electricity: the largest AI facilities can exceed 1 gigawatt of load. Building network connections and procuring turbines can take years; in many cases capacity is already booked through 2029. That means even with funding, new data centres cannot be brought online at desired scale if the power grid and network infrastructure cannot keep pace.
Geopolitics raises the stakes
Chips have become strategic. On March 19 in the United States, three people were charged with moving servers containing AI chips to China in apparent violation of export controls, involving roughly $2.5 billion in equipment. Such cases highlight that states are tightly monitoring technology flows because access to advanced hardware directly affects computing power and competitive position.
What will decide the race?
Given current conditions, the AI race will be decided less by who tells the most compelling story and more by who secures timely access to manufacturing capacity, memory, optical components, energy and packaging. The story today is less science-fiction breakthrough and more a sweat-and-industry struggle: beneath the demos and marketing, supply chains and infrastructure realities will determine who can actually scale AI in time.
Author: Kovács Sándor — SME strategy consultant, writer and speaker.
