The Long Shadow of a GPU: Second and Third-Order Effects of the AI Compute Build-Out

The Industrial Realignment of the AI Economy

For three years, the global conversation surrounding artificial intelligence has remained almost entirely concentrated on the abstract software layers at the top of the technology stack. Public attention has focused on frontier foundational models, autonomous agents, and conversational interfaces designed to replace or augment cognitive labour. However, the more consequential narrative of this technological era is unfolding far beneath the software layer. The build-out of AI compute has become, in less than 36 months, the largest industrial capital cycle since the post-war reconstruction of Europe. Hyperscalers are on pace to spend $620 billion on AI infrastructure in 2026 alone. Goldman Sachs expects AI to drive a 165% increase in data centre power demand by 2030.

Concentrated capital of such magnitude doesn't sit politely inside the tech sector. It ripples. It pulls memory wafers off smartphone production lines, transformers off utility queues, and copper off renewable energy projects.

Consequently, the most interesting story in the AI economy right now isn't being told from Cupertino or Santa Clara. Instead, they are being driven by a fine ceramics factory in Kitakyushu, Japan, an amino-acid chemical plant outside Kawasaki, Japan, an electrical-steel mill in Butler, Pennsylvania, and a commercial front-end semiconductor fabrication facility in Dholera, India.

The identified physical sites represent the foundational bottlenecks of the physical infrastructure that turns raw electricity into digital intelligence.
‍

First-order Impact: the Memory Squeeze

The most immediate and severe bottleneck of this infrastructure build-out is concentrated in the global memory chip market. The mechanics are simple. For example, an Nvidia B300 graphics processing unit (GPU) requires eight stacks of High-Bandwidth Memory (HBM), with each stack containing twelve dynamic random-access memory (DRAM) dies, resulting in a total of 96 DRAM dies per single GPU. Producing one bit of HBM consumes roughly three times the wafer capacity needed to produce one bit of conventional DDR5 memory. Furthermore, HBM sells for $60-100 per module versus $5-10 for an equivalent capacity of conventional DRAM. For the three companies that control 95% of global DRAM supply: Samsung, SK Hynix, and Micron, the math is unambiguous: every wafer that can be diverted to HBM, will be.

The result is the largest memory supply shock in 15 years. According to quarterly industry surveys compiled by TrendForce, DRAM contract prices rose 58-63% quarter-on-quarter in Q2 2026. NAND flash rose 70-75% in the same window, marking the steepest quarterly increase in a decade. One terabit of TLC NAND went from roughly $4.80 in July 2025 to about $10.70 by year-end. Enterprise SSDs are up 53-58% in a single quarter. In 2026, AI has become a major driver of the global memory shortage. OpenAI has reportedly secured agreements for up to 900,000 DRAM wafers a month for Stargate, potentially equivalent to around 40% of global DRAM output at full scale.

The current situation differs from the cyclical memory shortages of 2017 or 2021 because it isn't cyclical. It's a structural, possibly permanent, reallocation of the world's silicon wafer capacity. Industry executives, including the leadership of Phison and Silicon Motion, have noted that manufacturing capacity across DRAM, HBM, NAND, and traditional hard disk drives is entirely committed, with no meaningful supply relief projected until 2028.

The historical analogy isn't the dot-com bubble or the 1990s fiber boom. It's the 1970s oil shock. In both cases, a single industry's demand curve became steep enough to reprice an entire commodity for every other industry that used it. Just as steelmakers, airlines, and supermarkets discovered in 1973 that they were now competing for the same barrel of oil as the petrochemical industry, smartphone makers, automakers, and white-goods manufacturers are now discovering that they are competing for the same wafer as NVIDIA.
‍

Second-order Impact: the consumer electronics hit

Structural reallocation of wafer capacity has had an immediate impact on consumer electronics, raising component costs and forcing manufacturers to alter device designs.

International Data Corporation (IDC) now expects 2026 global smartphone sales to fall 12.9% to 1.12 billion units, marking the lowest sales volume in more than a decade. Average smartphone selling prices are projected to rise 14% to a record $523, and manufacturers will no longer be able to produce phones that retail below $100. Xiaomi's CFO has publicly flagged that DRAM costs per phone could rise ~25% in its 2026 model year. To preserve margins without passing the entire cost on to consumers, the company is redesigning baseline configurations, replacing older "8+128GB" storage setups with "4+256GB" storage. Other Android OEMs are doing the same.

IDC also expects the global PC market to shrink 5-9% in 2026, depending on severity, with average selling prices up 6-8% even in the moderate case. Gartner has begun forecasting that low-margin entry-level laptops under $500 will become financially unviable within two years. Dell, Lenovo, HP, Acer, ASUS have all warned clients of 15-20% hardware price hikes.

The most uncomfortable second-order consequence in price-sensitive markets like India is the cost structure of entry-level consumer electronics that is being reshaped in real time. The next 18-24 months are likely to favour brands that can manage component cost volatility well, hold price points through engineering trade-offs, and avoid passing the entire memory-cost shock through to end consumers.
‍

Third-Order Winners: AI beneficiaries nobody initially called AI stocks

The current capital-intensive cycle draws attention to highly specialized, capital-intensive manufacturing processes that are difficult to replicate. The global AI hardware supply chain relies on several specialized micro-monopolies, each controlling a critical step in the physical production process.

Advanced Wafer Handling and Substrate Insulation: Toto and Ajinomoto

Both stories went viral in February 2026 after the London-based activist investor Palliser Capital published a presentation arguing that Toto (TYO: 5332), Japan's best-known toilet maker, was one of the most undervalued AI beneficiaries on the planet. The stock jumped nearly 30% in a month, and again in May after Toto announced ¥30 billion ($188 million) of new investment in semiconductor ceramics through fiscal year 2028. Toto is the world’s second-largest manufacturer of electrostatic chucks (ESCs), precision ceramic pedestals that use electrostatic force to hold silicon wafers flat during high-aspect-ratio plasma etching. In advanced 3D NAND manufacturing, these chucks are essential for cryogenic etching tools and are projected to deliver 55% of group operating profit for the fiscal year ending March 2026. In profit terms, Toto’s globally famous toilet-and-bathroom business is now the smaller half of the company.

Ajinomoto, the Japanese company that invented MSG, is the other set-piece in this story. Leveraging its expertise in amino-acid chemistry, the company developed Ajinomoto Build-up Film (ABF), which is an ultra-fine, highly insulating epoxy-resin film. Today, ABF holds a 95%-98% market share in the advanced packaging substrates used by Nvidia, AMD, and Intel. Without ABF to insulate the microscopic copper circuits within these substrates, the high-speed interconnects would fail due to signal leakage and thermal breakdown. The market still values Ajinomoto primarily as a food company; the AI-exposed segment, valued at peer semiconductor multiples, may be worth nearly the entire current market cap on its own.
‍

The under-cited chokepoints

A few more stories that deserve more attention than they've gotten:

Cleveland-Cliffs (US, NYSE: CLF) is the only domestic producer of grain-oriented electrical steel (GOES) in the United States: the specialty silicon-steel used to make the magnetic cores of large power transformers. Without GOES, you cannot build a transformer; without transformers, you cannot connect a data centre to the grid. Demand for generator step-up transformers grew 274% between 2019 and 2025. Lead times for high-capacity units have stretched to four years. Hitachi Energy is investing over $1 billion in a new US plant scheduled for 2028; Siemens Energy is committing $421 million to a new factory in Charlotte. None of that supply comes online soon. And every transformer manufacturer drawing on domestic steel draws from a single source.

The downstream side of the same problem is just as interesting, and is where a new class of entrants is emerging. The incumbent transformer manufacturers: GE Vernova, Siemens Energy, Hitachi Energy, Eaton, have 18-36 month lead times and pricing power that is now untenable for the pace of data centre buildout. That gap is being filled by asset-light, India-leveraged entrants like Ayr Energy, which uses excess capacity at established Indian contract manufacturers to deliver custom-engineered transformers and switchgear into the US grid and data centre market at a fraction of incumbent lead times. It's a category where Indian manufacturing genuinely becomes part of the global AI supply chain, not just a downstream consumer of it.

Siemens Energy (Germany, ETR: ENR) is Europe's largest AI-theme winner by a wide margin, and almost nobody outside European industrials desks is framing it that way. The stock is up over 1,600% from its 2023 lows and 138% in 2026 alone, on the back of grid technology and gas-turbine demand from data centre customers. The CEO Christian Bruch has openly attributed the upgrade to AI data centre orders. BofA and JPMorgan are both calling it the top European industrials pick of 2026, with price targets of €170 and €160 respectively. It is, in practice, one of the most direct AI infrastructure plays in the world.

Asia Vital Components (Taiwan, TPE: 3017) is the dominant supplier of thermal management hardware like fans, heat sinks, 3D vapor chambers, liquid-cooling modules like in-rack/in-row coolant distribution units, copper cold plates, etc to Foxconn and to various hyperscalers. As server rack densities rise from 15kW to over 100kW, conventional air-cooling systems are reaching their physical limits, forcing a transition to liquid cooling. By leveraging this relationship, Foxconn has transitioned from a consumer electronics assembler into a high-value AI rack integrator. Taiwan supplies roughly 90% of global AI servers, and by 2025, AI and cloud were over 40% of Foxconn's revenue mix - filing away from the sole positioning of being Apple's manufacturer.

AI compute isn't really being bottlenecked by GPUs anymore. It's being bottlenecked by various other companies holding a near-monopoly on something microscopic - or boringly massive - that the entire stack depends on.
‍

Fourth-order Impact: copper, transformers, water, land

Beyond chips and components, the sheer physical scale of AI data centres is clashing with real-world resource constraints, turning electricity, copper, water, and land into highly contested commodities.

Copper.

A conventional data centre uses 5,000-15,000 tonnes of copper. A hyperscale AI data centre built for NVIDIA HGX systems can use up to 50,000 tonnes per facility. S&P Global expects AI and electrification together to drive a copper demand surge that the mining industry has not been investing to meet. Freeport-McMoRan, the largest US copper producer, has become a quiet AI proxy. As AI server racks move toward liquid cooling, which requires copper cold plates on every chip, the per-server copper intensity is rising further.

Power transformers.

‍The chokepoint becomes existential here. There is now a serious scenario where data centres that have been approved, built, and equipped with GPUs cannot turn on because the transformer to connect them to the grid hasn't arrived, as the Cleveland-Cliffs story above hints at. Sam Altman has talked publicly about OpenAI needing 250 gigawatts by 2030, more than India's entire installed grid capacity, for a single company. Virginia's grid operator has stated openly that some already-approved data centres may never come online because the power infrastructure cannot be built fast enough. McKinsey forecasts 100 GW of new global data centre capacity coming online between 2026 and 2030; whether even half of that actually gets energised is genuinely uncertain.

Diesel and natural gas.

Because grid connections are slow, data centres are increasingly being built behind-the-meter with on-site gas generation. Caterpillar, GE Vernova, and Cummins have all become AI-adjacent industrial stocks. Bloom Energy and FuelCell Energy have become AI-adjacent fuel cell stocks. Constellation Energy signed a 20-year power purchase agreement with Microsoft to restart Three Mile Island. Vistra Energy signed a 20-year PPA with Meta for all 2,176 MW of its Perry and Davis-Besse nuclear plants. Uranium miners: Cameco, Energy Fuels, Centrus, are trading like AI stocks.

Water and land.

AI data centres use millions of gallons of water per day for evaporative cooling. Hyperscale facilities increasingly occupy hundreds of thousands of square feet of industrial real estate. The McKinsey figure above on 100 GW of new capacity translates to roughly $1.2 trillion in real estate asset value and another $1-2 trillion in IT equipment fit-out - most of it bottlenecked by exactly the input constraints described above.

The deeper point is that AI has reached a scale where it is no longer a software industry. It is a heavy industry. It is closer in physical character to steel or refining than to SaaS. And the rents in heavy industry historically accrue to whoever has secured the inputs first.
‍

What this means for India

India sits in an unusual position in this story. It is simultaneously one of the world's most price-sensitive consumer device markets and one of the most strategically positioned countries to build the next layer of the AI stack. India could become increasingly important across the physical and operational layers of AI infrastructure.

1. Semiconductors and compute

Micron's HBM packaging facility in Sanand, Gujarat is the first project cleared under the India Semiconductor Mission that began commercial production in February 2026. It is the first time India has been integrated into the global HBM supply chain. Tata Electronics's 300-mm fab in Dholera, backed by an $11 billion investment and a partnership with ASML signed in May 2026, is targeted at 28-110nm nodes that are useful for industrial, automotive, and edge AI chips. The opportunity is considerably broader than leading-edge chips and includes packaging, specialty components, industrial chips and electronics manufacturing.

India is also taking a different approach to compute. MeitY has subsidised access, onboarding roughly 38,000 GPUs into a national compute facility, making them available to startups, researchers, and MSMEs at approximately ₹65 per GPU-hour, against global cloud rates of $2-3 per hour. This is a different bet from the hyperscaler model. The approach resembles India's broader digital infrastructure strategy.‍

2. The data centre build-out

Roughly 30 large data centre projects were announced in India between March 2025 and April 2026, adding 3.5 GW of planned capacity. Andhra Pradesh and Telangana alone account for over 2 GW, anchored by AI-specific mega-campuses in Visakhapatnam and Hyderabad. The L&T-NVIDIA sovereign AI factory in Chennai, the AdaniConneX expansions, the Yotta and CtrlS builds are all early bets that India will be a meaningful inference geography. Deloitte estimates India will need an additional 45-50 million sq. ft. of data centre real estate and 40-45 TWh of incremental power by 2030.

3. AI is becoming an industrial opportunity

The more interesting implication is that India's AI trade may be more industrial than technological. Every new data centre requires power, transformers, cables, switchgear, backup generation, cooling and connectivity. As AI workloads increase power density, these requirements become even more significant. Hitachi Energy India surged 12.5% in a single session in February 2026 on AI-infrastructure orders. Apar Industries rose 40% in a month on its cables-and-conductors exposure to AI data centre buildout. Netweb Technologies has an order book on track to roughly double over 30 months on AI-related orders. Cummins India, ABB India, L&T, and Voltas are all picking up AI-data-centre-linked orders in size.
‍

A closing thought

Every major industrial cycle has followed a similar arc. The railway mania of the 1840s built the network that powered the next century of industrialisation, even as many of the companies that built it went bankrupt. The telecom boom of the 1990s produced an enormous fibre build-out; many of its builders disappeared, but the infrastructure survived and became the backbone of the internet. As Ben Thompson has put it, the builders of the railroads went bust but left behind railroads; the electricity companies went bust but left behind electricity; the dot-com companies went bust but left behind fibre.

AI is beginning to look like the next version of this cycle. Not every dollar of the trillion-dollar AI capex wave will earn an attractive return. Some infrastructure will be overbuilt, some assets will become stranded, and some will eventually look like the “dark fibre” of the next generation. The more important question is not whether the cycle eventually breaks. It is who owns the infrastructure that survives it.

The most durable winners are likely to be the Totos: companies that were doing something difficult and essential long before anyone called it an AI business, and will still be doing it after the AI label disappears. Companies that own a genuine chokepoint, have capabilities that are difficult to replicate, and sit deep enough in the supply chain that customers cannot easily substitute them away.

For India, this creates a particularly interesting opportunity. The country does not need to win the frontier-model race to capture a meaningful share of the AI economy. It needs to identify the chokepoints where it can build globally competitive capabilities from specialty materials, semiconductor packaging to power equipment, cooling, cables, data infrastructure and embedded intelligence. Some of India's future Totos may be companies that today look decidedly non-AI: a specialty manufacturer that becomes indispensable to global electronics, an industrial technology company whose systems end up inside the next generation of machines, or an infrastructure business that quietly becomes critical to hyperscalers.

India has the ingredients: engineering talent, a large domestic market, competitive manufacturing economics and a policy push that is increasingly focused on building foundational capacity rather than chasing headlines.

The question is whether it can turn those advantages into global chokepoints before the rest of the world moves on.

‍

DISCLAIMER

<sub>Details of all the funds managed by 3one4 capital can be found here: https://www.3one4capital.com/aif-registration-details

For informational purposes only. This content does not constitute an offer to sell or a solicitation of an offer to buy units of any fund. Investments in AIFs involve market risks.</sub>

_{The views expressed herein are those of the author as of the publication date and are subject to change without notice. Neither the author nor any of the entities under the 3one4 Capital Group have any obligation to update the content. This publications are for informational and educational purposes only and should not be construed as providing any advisory service (including financial, regulatory, or legal). It does not constitute an offer to sell or a solicitation to buy any securities or related financial instruments in any jurisdiction. Readers should perform their own due diligence and consult with relevant advisors before taking any decisions. Any reliance on the information herein is at the reader's own risk, and 3one4 Capital Group assumes no liability for any such reliance.Certain information is based on third-party sources believed to be reliable, but neither the author nor 3one4 Capital Group guarantees its accuracy, recency or completeness. There has been no independent verification of such information or the assumptions on which such information is based, unless expressly mentioned otherwise. References to specific companies, securities, or investment strategies are not endorsements. Unauthorized reproduction, distribution, or use of this document, in whole or in part, is prohibited without prior written consent from the author and/or the 3one4 Capital Group.}

‍