Scholars and publications that influenced me
A handful of brilliant minds - Kevin Kendall, Manoj K. Chaudhury, Kenneth R. Shull, Costantino Creton, Alan N. Gent, Steven Abbott, and Zhigang Suo - have really shaped the way I think about adhesion science and soft materials. If you're a student or a professional looking to get a better grip on how adhesives and coatings actually work (and how to design and test them), the articles below are solid go-tos.
Scaling my insights
During grad school, I studied the mechanics of adhesives and soft materials. I really enjoyed understanding why things stick, don't stick, slip, and break. That experience helped me develop a framework to understand adhesion. After moving to industry, that framework deepened by working with many different materials and solving real problems customers face.
Since moving to industry, I realized that I needed a framework to understand markets. Looking back, I never imagined I'd enjoy reading 10-Qs and 10-Ks. I never thought I'd look forward to listening to earnings calls. Yet these sources offer insight into companies and industries that you can't find anywhere else in today's noise-filled world.
Lately, I've been deep-diving into AI infrastructure. Direct CapEx for U.S. data center build-out is estimated a $1.8 trillion between 2024 and 2030. That does not include grid modernization and new power plant construction. For perspective, the U.S. government spent an inflation-adjusted $5 trillion on World War II from 1941 to 1945. Explosive investment.
Take AI workloads as an example. They're scaling up by orders of magnitude every year. The amount of tokens generated in a given period now is > 10x what was the same period the year before. And it's expected to keep scaling like this. This massive scale has turned into a key profit driver for big tech hyperscalers. But it creates a serious challenge: power management. Why? Because AI data centers swing massively in power demand - sometimes within milliseconds. If those swings aren't smoothed out, blackout risks rise. Meanwhile, the solar power duck curve keeps getting worse year after year, adding complexity to grid stability.
On top of that, full-stack AI cluster management is getting critical especially for enterprises wanting to avoid full reliance on cloud providers. The goal is a simple software abstraction layer over complex hardware layer made up of a bunch of IT equipment, cooling and power modules. This lets non-technical users in enterprises and governments operate AI infrastructure effectively. But this requires new skills beyond traditional enterprise IT setup. Getting it right is key to really squeezing out ROI as AI moves from hype to mainstream. The big question: who can make AI adoption seamless and profitable for all, not just push its frontier?
AI and data sovereignty have become critical priorities for both private and public sectors. Power management and the architecture of AI infrastructure are the backbone of these priorities - on par with compute, networking, and advanced cooling solutions. If we don't solve these challenges, next-generation, gigawatt-scale data centers simply will not be viable.
Remember the 2021 winter blackouts in Texas? Today, Texas faces more than 100 GW of new load requests-mostly from data centers-trying to connect to its power grid. At the same time, over 400 GW of new generation projects await grid interconnection. For perspective, U.S. peak load typically ranges between 700 and 800 GW. Texas' reserve margins are expected to turn negative by 2028, meaning demand will exceed available supply. Meanwhile, Nvidia is building an AI ecosystem across Arizona and Texas. with chips made at TSMC's Arizona fabs and AI supercomputers assembled in Texas. The ecosystem will attract even more energy-intensive projects. Notably while AI's carbon footprint gets attention, the water footprint remains under-discussed despite being equally critical.
Fracking in the Permian Basin demands huge volumes of water around 4 barrels of water for every barrel of oil produced. The water-to-oil ratio increases as wells age, sometimes reaching up to 10:1 or more in sub-basins. Despite oil price pressures and tariffs tempering new rig investments, oil production continues hitting new records by milking existing rigs. Oil and gas companies have traditionally injected produced water into underground disposal wells. But Texas regulators are cracking down on this practice amid rising seismic activity linked to wastewater injection and environmental safety concerns. Many disposal wells are approaching capacity limits. Handling produced water directly by O&G companies have been eroding their profits, and it will only get worse. The produced water disposal cost doubles every 5 years, which is equivalent to a CAGR of 15%.
Yet, this challenges also creates opportunity. Landowners with surface rights controlling large, unutilized or underutilized pore spaces may profit by offering disposal capacity. Meanwhile, companies innovating in produced water treatment and recycling see growing demand. Recycling rates in the Permian are rising rapidly. In 2023, more produced water was used for fracking completions than fresh or brackish water sources. Produced water remains a significant liability for oil and gas companies. Not managing it well risks costly regulatory penalties and operational bottlenecks. Money flows where the pain is. Companies like Cintas show how businesses can profit by providing essential services that help others avoid compliance risks and focus on core operations.
Beyond data centers, industrial electrification is accelerating to cut CO2 emissions and enable smart manufacturing automation. The industrial sector consumes about 35% of U.S. energy, but only 10% is electricity while 80% comes from hydrocarbon fuels. Oils and natural gas remain more efficient and cost-effective for achieving the very high temperatures required in chemical processes for steel, cement, glass, plastics, and semiconductor production. Leading chemical companies are piloting electrically heated steam crackers. For smart manufacturing, where 5G, sensors, and actuators control precise and rapid reaction processes, electricity is essential. But here's the catch: where will this electricity come from? PJM Interconnection, the largest power grid operator in the U.S., conducts annual capacity auctions to ensure sufficient future power supply. Auction clearing prices have surged dramatically over 3 years: $29/MW day (2024/2025), $270/MW day (2025/2026), $330/MW day (2026/2027). This represents over a 1,000% increase, driven primarily by booming data center demand.
To rapidly respond to energy demands, natural gas power plants are the only viable option for the next decade. They're easily scalable, can come online quickly (< 1 year), and benefit from very cheap natural gas. But we need a lot more energy for AI and EVs. That's why in May 2025 Trump signed executive orders to revive the nuclear industry by aiming to build 300 nuclear plants with 400 GW capacity by 2050. Easier said than done. Most U.S. nuclear plants were built between 1970 and 1990. Since the Three Mile Island incident in 1979, only two entirely new reactors have had their construction initiated and completed. As a result, the U.S. nuclear construction capability and supply chains have significantly weakened, though companies like GE Vernova still have strong capabilities in nuclear plant design and engineering.
The bottom line is that the U.S. can't do it alone and is teaming up with global powerhouses like Korea. With onshore manufacturing in focus, U.S. companies making radiation-shielding steels, seals, pressure vessels, steam generators, heat exchangers for both gigawatt-scale traditional nuclear plants and SMR are set for a boom. Interestingly, some of these companies also supply heat exchangers and turbines to the Navy, which has recently been investing heavily in shipbuilding and its nuclear strategy like Columbia and Virginia Class submarines.
Okay, enough about AI infrastructure. What's the real untapped money-making opportunity for big techs betting big on AI? It's not just about chatbots. The real value lies at the endpoint of the AI value stream - 'AI for the physical world'. Imagine real-time AI data analytics connected to billions of sensors, working like "meters" in taxis. That's where real cash flow comes from, powered by token-based billing systems.
Big tech is already moving in this direction by partnering with legacy manufacturers. FDA-approved AI medical devices are rising. AI-operated drones are proving their worth, and there's a huge push to invest in data links, edge computing, and embedded software to keep pace with rapidly evolving warfare tactics. Data analytics and software-defined weapons are being woven directly into military operations. Defense as a Service is emerging as a viable business model for defense contractors, helping to smooth out lumpy funding cycles.
Yet, as of early 2025, AI adoption in logistics, warehousing, and manufacturing remains below the 7% industry average. What bottlenecks slow broader adoption? The White House AI Action Plan points to three main issues: lack of understanding of AI, a complex regulatory landscape, and lack of clear governance. Notably, the plan emphasizes a shift towards a "try-first" culture across American industry - a major change from the traditional tone of regulations and bureaucracy.
AI infrastructure leaders see us entering the second wave of adoption, AI inferencing, focused on deploying verticalized AI at the enterprise level. This follows the first wave, AI training, driven by foundational models from big tech using lots of NVIDIA A100 and H100/H200 chips. AI training is a cost, but inference is revenue. It means going forward they'll push to embed inference-based platforms (that run on Blackwell chips) as recurring revenue streams within enterprises' operating expenses, essentially making the platform super sticky within their operations.
Until late 2024, investors focused on chip makers, hyperscalers, gas turbine/HVAC producers, data center builders, and independent power producers (IPP). During this time, The AI = Scaling framework became widely accepted in AI training. This approach demands lots of GPUs, electricity, cooling, and new space. Valuations surged, reflecting what many (myself included) see as a bubble in the AI market, though it's hard to dismiss some intrinsic value.
But people are now talking about the limits of scaling, and as we pivot to AI inferencing, having huge infrastructure becomes less critical and can actually slow down adoption. If society doesn't adopt AI widely, hyperscalers (Amazon, Google, Microsoft) and model developers (OpenAI, Anthropic) won't be able to justify their massive investments. Their real opportunity is to sell proprietary APIs to enterprises and generate recurring revenue - a frontier they are eager to break through.
Basically, they are now focused on operating AI factories that sell cash-printing products (APIs). This is the only way to vindicate the humungous investment. The key market drivers are no longer limited to procuring AI hardware and building AI data centers (2023-2025), a trend sparked by ChatGPT, although this trend will continue. The next phase (2025+) is achieving maximum operational efficiency in these AI factories.
How can this be achieved? Let's do some thought experiments, assuming no physical/engineering limits. Maximum ROI and total market dominance will come from 100% compute utilization, zero downtime, infinite bandwidth, and zero latency. Those who can meaningfully converge these elements will ride strong secular tailwinds. Unless an AI winter comes (not yet) or the AI bubble pops (possibly), their valuations, still flying under the radar, will skyrocket over the next few years, just like what happened with Celestica, Credo, Modine, Vertiv, Sterling Infrastructure, and Vistra between 2023 and 2025.
Looking ahead, new areas that are getting noticed or that I expect will soon include:
That said, I'm curious how the tariffs on steel, aluminum, and copper products will impact market dynamics. These three materials are the bread and butter of AI infrastructure. Transformers, which are made from many of these elements, are in severe shortage right now due to the rapid growth of data centers. About 80% of transformers are imported, and they are also subject to tariffs. Because these tariffs are imposed under Section 232 for national security reasons, it's tough for courts to overturn them. The steel and aluminum tariffs that first came in back in 2018 during President Trump's admin have stuck around through his first term and haven't been lifted by the Biden admin. But things were quite different then. U.S. inflation was low, and borrowing costs for businesses were much cheaper. Now, in 2025, we're looking at a totally different economic landscape.
According to recent earning calls by leading communication companies, after a downcycle in 2023 and 2024, wireless service providers are ramping up investment to renovate infrastructure, but rural areas may be put a a disadvantage. The Biden admin's BEAM (Broadband Equity, Access, and Deployment) program strongly favored a 'fiber-first' approach for accessing high-speed internet, but the policy has been significantly updated to replace the fiber-first stance with a 'technology-neutral' approach like satellite communications. Installing fiber optics networks is more expensive in rural areas than in urban centers, and rising component prices will further increase costs.
That being said, 5G networks are being pursued as effective rural broadband solutions. The One Big Beautiful Bill (OBBB) Act will speed up 5G and private network rollouts. The Trump admin restored the FCC's authority to auction spectrum, which had expired in 2023, and reinstated 100% bonus depreciation, allowing MNOs and enterprises to write off the full cost of 5G network equipment in the year of purchase. Starting in 2026, FCC will auction substantial amounts of mid-band and low-band spectrum over the next 8 years, which is key to enabling a nationwide 5G network. This is a huge catalyst for 5G network equipment providers.
Besides AI infrastructure, I've been closely following developments in quantum sensing, and more broadly, in Positioning, Navigation, and Timing (PNT), especially as 5G/6G frequency bands overlap or sit right next to military bands like the lower three GHz band used for missile defense. Quantum sensing can complement or serve an alternative to traditional GPS. Unlike GPS, it is anti-jamming, anti-spoofing, and capable of detecting extremely weak signals. Most of the buzz is around quantum computing and startups that lack serious technology. PNT doesn't sound as exciting as AI and robotics, but GPS (an "invisible utility") underpins ground/aviation/maritime transport, financial transactions, power grid operations, agriculture, LEO satellite synchronization, and even AI and robotics. Disruption is already real. 85% of flights in Estonia experience GPS disruption because of Russia's electronic warfare against Ukraine. It is even affecting LEO satellites as they pass over Ukraine. In late 2024, the DoD raised the budget limit to use commercial LEO constellations from $900 million to $13 billion. I'm sure the DoD won't be pleased if the same kind of disruption starts hitting U.S. assets.
The Achilles' heel of drone warfare is the lack of signal resilience against EMI. PNT is now slowly coming to the forefront of investors' attention for enabling two key initiatives: Golden Dome and LEO satellite proliferation. Detecting hypersonic missiles and managing satellite and even emerging eVTOLs traffic requires high-precision time and frequency measurement on the picosecond scale, or less than a billionth of a second. Quantum sensors are (quietly) being field-tested in the defense and space sectors. 2025 marks an inflection point, with the focus now on ruggedized quantum sensors agnostic to military platform and environment beyond the static lab setting. It's exciting to see how quickly the space ecosystem is evolving.
Speaking of LEO satellite proliferation, China is on track to complete the first phase of the SpaceSail Project by the end of 2025, with a target of launching 15,000 LEO satellites by 2030 as a direct competitor to SpaceX's Starlink. The rivalry is no longer limited to AI and semiconductors. Space has become a new arena in the escalating U.S.-China competition. The space industry is characterized by a 'first come, first served' dynamic and is key to the global digital order - no matter what the ultimate goal may be (which is kind of scary). Personally, I enjoy analyzing the space industry's value chain, from satellite manufacturers to operators, ground equipment suppliers, service providers (MNOs), and end users. It's quite vibrant; for example, advanced terminals for in-flight connectivity (IFC) are line-fitted or retrofitted on lots of commercial and business aircraft annually.
This is my own half-baked philosophy, but I think that the common thread running through themes like AI, robotics, advanced computing, communications, missile defense systems, space technology, and quantum science is PNT. It's actually the biggest single point of failure in all of these fields. In the era of software-defined everything (SDx/SDE) across the commercial, civil, and defense sectors, a tiny misstep like a split second error or PNT interference can have a huge impact on machine efficiency and human lives.
Whoever productizes hardware and software that transport electrons and photons faster, denser, with less loss, lower power, lower noise, tighter sync will ride this AI era. A perfect example is Co-Packaged Optics (CPO), now integrated into the Blackwell platform to replace power-hungry Digital Signal Processor (DSP). On top of that, Nvidia is reportedly exploring a new packaging architecture called CoWoP, set to replace the current CoWoS used across all GPUs, including Blackwell. This approach removes the substrate (S), mounting the chip interposer (CoW) directly onto the PCB (P). Easier said than done. It puts immense pressure on PCB value chains. Resonac, a key player, has already formed a consortium with leading semi companies to address it.
Lord Kelvin once said, "If you can't measure it, you can't improve it." Look at the LLM benchmarking wars, where leaders clash and leapfrog each other almost weekly or monthly. The ranks basically reflect who has the most advanced system to measure and improve the flow of electrons and photons, the fundamental carriers of data.
In the age of AI inferencing, I think the real frontier may be the network - the layer that governs how electrons and photons flow at every scale - server, rack, cluster, data center fabric (Broadcom, Arista, Nvidia, Credo, Coherent, Marvell, Lumentum, MaxLinear), inter data center (Ciena), national backbone (AT&T, Verizon, Equinix, Lumen), intercontinental, space (SpaceX, Amazon, AST SpaceMobile, Viasat), and ultimately edge (Akamai, Cloudflare). From backhaul to fronthaul, from first mile to last mile, numerous competitors are trying to eat each other's lunch within this fragmented network market.
Unless you're interested in technology purely for its own sake, staying informed about macroeconomic trends and geopolitics is essential to understanding how the technology landscape evolves. Investment volumes shift dynamically, and a vertical that was once second-tier in the global market can become first-tier, and the reverse can also happen.
For example, while jamming and spoofing pose real threats to commercial airlines, the development and deployment of quantum sensors depend largely on federal funding for defense applications. Monitoring the defense budget and understanding how funds are allocated across different focus areas is critical. I enjoy reading abstracts of awardees in the Small Business Innovation Research (SBIR) program to learn about early-stage R&D activities by small businesses. The DoD is an attractive target market, as it adopts new technologies quickly and is flexible about procedures or company size as long as performance is validated.
In FY 2026, sensing capabilities and satellite communications are the top priorities in support of the Golden Dome initiative. Photonics and PNT technologies, which I mentioned earlier, are central to these efforts. Also, space-grade alloys and composites are needed to intercept missiles and reduce payload weight. Just as critical is the systems engineering that ties everything together. It's like owning a stack of high-end Nvidia GPUs makes no sense if you don't have the right abstraction layer to direct and route the flow of data. It's quite fun to learn all these advances (though it is kind of scary).
Lastly, a lot of folks are talking about market bubbles, potential collapses, and recessions. The current trend seems to be that whenever (minor) corrections happen, massive dip-buying follows. Many investors don't seem scared of corrections anymore. They see them as chances to scoop up bargains. The U.S. has been on an easy money steroid shot since 2008, driven by multiple rounds of quantitative easing (QE), years of ultra-low interest rates, and lavish fiscal policy. Even after 3 years of quantitative tightening (QT) since mid-2022, there's still plenty of money (M2) circulating around, which has kept the market artificially propped up. The Fed's balance sheet shows a correlation with the S&P500.
Are market valuations stretched? Yes. Will the U.S. economy hit a recession? Recent macro data aren't pointing in that direction yet. BUT, major corrections triggered by disruptive policy changes and geopolitical events could cause a recession, rather than by themselves. That's just my armchair take, with all the risk of being totally wrong and facing the humiliation! I'm not a trained economist or a professional market analyst, just someone who thinks it's worth building a rough framework even as a layperson.
Here are some useful links to help follow macroeconomic trends for laypeople like me.
My framework for understanding market trends is built on irreversible, inevitable forces, independent of contentious AI bubbles.
A handful of brilliant minds - Kevin Kendall, Manoj K. Chaudhury, Kenneth R. Shull, Costantino Creton, Alan N. Gent, Steven Abbott, and Zhigang Suo - have really shaped the way I think about adhesion science and soft materials. If you're a student or a professional looking to get a better grip on how adhesives and coatings actually work (and how to design and test them), the articles below are solid go-tos.
- Thin-film peeling-the elastic term, K. Kendall, J. Phys. D: Appl. Phys., 1975 (Role of an elastic deformation term on adhesion)
- Fracture mechanics and the adherence of viscoelastic bodies, D. Maugis and M. Barquins, J. Phys. D: Appl. Phys., 1978 (Validated the velocity dependence of the energy loss term across various loading geometries)
- The peeling of flexible laminates, A. J. Kinloch, C. C. Lau, and J. G. Williams, International Journal of Fracture, 1994 (Detailed factors for determining the adhesion energy in peel mode)
- Molecular Weight Effects in Chain Pullout, C. Creton, H. R. Brown, and K. R. Shull, Macromolecules, 1994 (The effect of molecular weights for bonding two layers with diblock copolymers and their chain pullout)
- Effect of interfacial slippage on viscoelastic adhesion, B. Zhang Newby and M. K. Chaudhury, Langmuir, 1997 (Friction matters in peel release of adhesive from coating)
- Effects of Methylation and Neutralization of Carboxylated Poly(n-butyl acrylate) on the Interfacial and Bulk Contributions to Adhesion, D. Ahn and K. R. Shull, Langmuir, 1998 (The effect of surface chemistry and bulk rheology on adhesion)
- Axisymmetric adhesion tests of soft materials, K. R. Shull, D. Ahn, W-L. Chen, C. M. Flanigan, and A. J. Crosby, Macromol. Chem. Phys., 1998 (Review/experimental paper on details of interfacial contributions to overall adhesion performance from the revised JKR approach that considers finite size effects)
- On Stickiness, C. Gay and L. Leibler, Physics Today, 1999 (A short yet insightful 5-page essay on the complexity of adhesion)
- Contact mechanics and the adhesion of soft solids, K. R. Shull, Mater. Sci. Eng. R Rep., 2002 (Comprehensive review paper on contact mechanics of soft materials)
- Adhesion between Immiscible Polymers Correlated with Interfacial Entanglements, P. J. Cole, R. F. Cook, and C. W. Macosko, Macromolecules, 2003 (The effect of molecular weights for bonding immiscible polymers through entanglements)
- Adhesion between a Viscoelastic Material and a Solid Surface, F. Saulnier, T. Ondarçuhu, A. Aradian, and E. Raphaël, Macromolecules, 2004 (Multiple time and length scales within the peel process)
- Soft and hard adhesion, J. Y. Chung and M. K. Chaudhury, J. Adhes., 2005 (Pull off an object from thin and thick coatings)
- Fracture and adhesion of soft materials: a review, C. Creton and M. Ciccotti, Rep. Prog. Phys., 2016 (Best review paper that provides a framework for thinking about designing and understanding elastomers, gels, and PSAs from the polymer engineering perspectives)
- Soft Materials by Design: Unconventional Polymer Networks Give Extreme Properties, X. Zhao, X. Chen, H. Yuk, S. Lin, and G. Parada, Chemical Reviews, 2021 (Must-read review paper for an aspiring soft matter engineer who wishes to adopt a framework for designing gels, elastomers, and adhesives)
Scaling my insights
During grad school, I studied the mechanics of adhesives and soft materials. I really enjoyed understanding why things stick, don't stick, slip, and break. That experience helped me develop a framework to understand adhesion. After moving to industry, that framework deepened by working with many different materials and solving real problems customers face.
Since moving to industry, I realized that I needed a framework to understand markets. Looking back, I never imagined I'd enjoy reading 10-Qs and 10-Ks. I never thought I'd look forward to listening to earnings calls. Yet these sources offer insight into companies and industries that you can't find anywhere else in today's noise-filled world.
Lately, I've been deep-diving into AI infrastructure. Direct CapEx for U.S. data center build-out is estimated a $1.8 trillion between 2024 and 2030. That does not include grid modernization and new power plant construction. For perspective, the U.S. government spent an inflation-adjusted $5 trillion on World War II from 1941 to 1945. Explosive investment.
Take AI workloads as an example. They're scaling up by orders of magnitude every year. The amount of tokens generated in a given period now is > 10x what was the same period the year before. And it's expected to keep scaling like this. This massive scale has turned into a key profit driver for big tech hyperscalers. But it creates a serious challenge: power management. Why? Because AI data centers swing massively in power demand - sometimes within milliseconds. If those swings aren't smoothed out, blackout risks rise. Meanwhile, the solar power duck curve keeps getting worse year after year, adding complexity to grid stability.
On top of that, full-stack AI cluster management is getting critical especially for enterprises wanting to avoid full reliance on cloud providers. The goal is a simple software abstraction layer over complex hardware layer made up of a bunch of IT equipment, cooling and power modules. This lets non-technical users in enterprises and governments operate AI infrastructure effectively. But this requires new skills beyond traditional enterprise IT setup. Getting it right is key to really squeezing out ROI as AI moves from hype to mainstream. The big question: who can make AI adoption seamless and profitable for all, not just push its frontier?
AI and data sovereignty have become critical priorities for both private and public sectors. Power management and the architecture of AI infrastructure are the backbone of these priorities - on par with compute, networking, and advanced cooling solutions. If we don't solve these challenges, next-generation, gigawatt-scale data centers simply will not be viable.
Remember the 2021 winter blackouts in Texas? Today, Texas faces more than 100 GW of new load requests-mostly from data centers-trying to connect to its power grid. At the same time, over 400 GW of new generation projects await grid interconnection. For perspective, U.S. peak load typically ranges between 700 and 800 GW. Texas' reserve margins are expected to turn negative by 2028, meaning demand will exceed available supply. Meanwhile, Nvidia is building an AI ecosystem across Arizona and Texas. with chips made at TSMC's Arizona fabs and AI supercomputers assembled in Texas. The ecosystem will attract even more energy-intensive projects. Notably while AI's carbon footprint gets attention, the water footprint remains under-discussed despite being equally critical.
Fracking in the Permian Basin demands huge volumes of water around 4 barrels of water for every barrel of oil produced. The water-to-oil ratio increases as wells age, sometimes reaching up to 10:1 or more in sub-basins. Despite oil price pressures and tariffs tempering new rig investments, oil production continues hitting new records by milking existing rigs. Oil and gas companies have traditionally injected produced water into underground disposal wells. But Texas regulators are cracking down on this practice amid rising seismic activity linked to wastewater injection and environmental safety concerns. Many disposal wells are approaching capacity limits. Handling produced water directly by O&G companies have been eroding their profits, and it will only get worse. The produced water disposal cost doubles every 5 years, which is equivalent to a CAGR of 15%.
Yet, this challenges also creates opportunity. Landowners with surface rights controlling large, unutilized or underutilized pore spaces may profit by offering disposal capacity. Meanwhile, companies innovating in produced water treatment and recycling see growing demand. Recycling rates in the Permian are rising rapidly. In 2023, more produced water was used for fracking completions than fresh or brackish water sources. Produced water remains a significant liability for oil and gas companies. Not managing it well risks costly regulatory penalties and operational bottlenecks. Money flows where the pain is. Companies like Cintas show how businesses can profit by providing essential services that help others avoid compliance risks and focus on core operations.
Beyond data centers, industrial electrification is accelerating to cut CO2 emissions and enable smart manufacturing automation. The industrial sector consumes about 35% of U.S. energy, but only 10% is electricity while 80% comes from hydrocarbon fuels. Oils and natural gas remain more efficient and cost-effective for achieving the very high temperatures required in chemical processes for steel, cement, glass, plastics, and semiconductor production. Leading chemical companies are piloting electrically heated steam crackers. For smart manufacturing, where 5G, sensors, and actuators control precise and rapid reaction processes, electricity is essential. But here's the catch: where will this electricity come from? PJM Interconnection, the largest power grid operator in the U.S., conducts annual capacity auctions to ensure sufficient future power supply. Auction clearing prices have surged dramatically over 3 years: $29/MW day (2024/2025), $270/MW day (2025/2026), $330/MW day (2026/2027). This represents over a 1,000% increase, driven primarily by booming data center demand.
To rapidly respond to energy demands, natural gas power plants are the only viable option for the next decade. They're easily scalable, can come online quickly (< 1 year), and benefit from very cheap natural gas. But we need a lot more energy for AI and EVs. That's why in May 2025 Trump signed executive orders to revive the nuclear industry by aiming to build 300 nuclear plants with 400 GW capacity by 2050. Easier said than done. Most U.S. nuclear plants were built between 1970 and 1990. Since the Three Mile Island incident in 1979, only two entirely new reactors have had their construction initiated and completed. As a result, the U.S. nuclear construction capability and supply chains have significantly weakened, though companies like GE Vernova still have strong capabilities in nuclear plant design and engineering.
The bottom line is that the U.S. can't do it alone and is teaming up with global powerhouses like Korea. With onshore manufacturing in focus, U.S. companies making radiation-shielding steels, seals, pressure vessels, steam generators, heat exchangers for both gigawatt-scale traditional nuclear plants and SMR are set for a boom. Interestingly, some of these companies also supply heat exchangers and turbines to the Navy, which has recently been investing heavily in shipbuilding and its nuclear strategy like Columbia and Virginia Class submarines.
Okay, enough about AI infrastructure. What's the real untapped money-making opportunity for big techs betting big on AI? It's not just about chatbots. The real value lies at the endpoint of the AI value stream - 'AI for the physical world'. Imagine real-time AI data analytics connected to billions of sensors, working like "meters" in taxis. That's where real cash flow comes from, powered by token-based billing systems.
Big tech is already moving in this direction by partnering with legacy manufacturers. FDA-approved AI medical devices are rising. AI-operated drones are proving their worth, and there's a huge push to invest in data links, edge computing, and embedded software to keep pace with rapidly evolving warfare tactics. Data analytics and software-defined weapons are being woven directly into military operations. Defense as a Service is emerging as a viable business model for defense contractors, helping to smooth out lumpy funding cycles.
Yet, as of early 2025, AI adoption in logistics, warehousing, and manufacturing remains below the 7% industry average. What bottlenecks slow broader adoption? The White House AI Action Plan points to three main issues: lack of understanding of AI, a complex regulatory landscape, and lack of clear governance. Notably, the plan emphasizes a shift towards a "try-first" culture across American industry - a major change from the traditional tone of regulations and bureaucracy.
AI infrastructure leaders see us entering the second wave of adoption, AI inferencing, focused on deploying verticalized AI at the enterprise level. This follows the first wave, AI training, driven by foundational models from big tech using lots of NVIDIA A100 and H100/H200 chips. AI training is a cost, but inference is revenue. It means going forward they'll push to embed inference-based platforms (that run on Blackwell chips) as recurring revenue streams within enterprises' operating expenses, essentially making the platform super sticky within their operations.
Until late 2024, investors focused on chip makers, hyperscalers, gas turbine/HVAC producers, data center builders, and independent power producers (IPP). During this time, The AI = Scaling framework became widely accepted in AI training. This approach demands lots of GPUs, electricity, cooling, and new space. Valuations surged, reflecting what many (myself included) see as a bubble in the AI market, though it's hard to dismiss some intrinsic value.
But people are now talking about the limits of scaling, and as we pivot to AI inferencing, having huge infrastructure becomes less critical and can actually slow down adoption. If society doesn't adopt AI widely, hyperscalers (Amazon, Google, Microsoft) and model developers (OpenAI, Anthropic) won't be able to justify their massive investments. Their real opportunity is to sell proprietary APIs to enterprises and generate recurring revenue - a frontier they are eager to break through.
Basically, they are now focused on operating AI factories that sell cash-printing products (APIs). This is the only way to vindicate the humungous investment. The key market drivers are no longer limited to procuring AI hardware and building AI data centers (2023-2025), a trend sparked by ChatGPT, although this trend will continue. The next phase (2025+) is achieving maximum operational efficiency in these AI factories.
How can this be achieved? Let's do some thought experiments, assuming no physical/engineering limits. Maximum ROI and total market dominance will come from 100% compute utilization, zero downtime, infinite bandwidth, and zero latency. Those who can meaningfully converge these elements will ride strong secular tailwinds. Unless an AI winter comes (not yet) or the AI bubble pops (possibly), their valuations, still flying under the radar, will skyrocket over the next few years, just like what happened with Celestica, Credo, Modine, Vertiv, Sterling Infrastructure, and Vistra between 2023 and 2025.
Looking ahead, new areas that are getting noticed or that I expect will soon include:
- Modular data centers serving edge locations (banks, law firms, healthcare providers, and telecoms) that need rapid AI deployment and future scalability with reasonable expense, but can't rely on hyperscalers due to security and regulatory limits. The concept itself isn't new; modular data centers saw limited use in the past but have gained momentum recently by solving a key AI infrastructure bottleneck: fast deployment. Major infrastructure providers are now introducing modular plug-and-play platforms for IT/cooling/power as faster, more flexible alternatives to traditional stick-built methods. Also, the industry perception is shifting: modular facilities are no longer seen merely as a way to augment traditional data centers, but as efficient prefab solutions for delivering very dense computing more. Shorter product cycles and the growing complexity of on-site builds make this shift even more compelling. Edge data centers are also becoming central to AI inference. Their role is analogous to Content Delivery Networks (CDN), which allows companies like Netflix, Amazon, and Meta to serve the majority of web traffic locally. Similarly, edge data centers expend their role beyond CDN to AI inference by caching copies of popular and long-tail AI models closer to users to reduce latency. The DoD is heavily investing in upgrading military equipment, with a primary emphasis on ruggedized High Performance Edge Computing (HPEC) systems designed to operate in harsh environments. Building on the momentum of modular data centers, Neocloud is an emerging concept for the same reasons above - flexibility, speed, and attractive pricing models. Neocloud platforms could play a key role in the 2nd wave of AI adoption, especially for small-to-medium enterprises and academic researchers who want to rapidly prototype and scale models but can't afford the high rental prices set by traditional cloud service providers.
- Low-emission microgrid power and chiller solutions that ease grid stress and enable edge data centers closer to crowded urban areas for enhanced latency. A leading power generator supplier recently said that the deficit in backup generators for data centers is over 5,000 units for 2026 alone. Along with transformers, backup generators are major concerns when it comes to lead times for data center developers and operators.
- Photonics to solve the memory wall and power efficiency challenges in AI servers. Recent trends are shifting away from monolithic servers towards heterogeneous fabrics based on scale-out disaggregated AI hardware, which offers tremendous flexibility in scaling capacity. Optical interconnects are the ultimate solution to make this architecture a reality. Copper interconnects are limited by huge amounts of heat caused by electrons colliding with copper atoms during massive data transfer. Beyond data centers, photonics is drawing enormous attention for its ability to collect more data, faster, and with less noise. It's already proving to be a key enabler for applications such as LEO satellite communications, military aiming systems, augmented reality, and biosensing for drug delivery. See how LiDAR that is the pinnacle of photonics is being rapidly adopted in the automotive, agriculture, and robotics industries.
- Telecommunication network to facilitate mass adoption of AI. Photonics in the above bullet point help handle AI workload capacity. But in broader data transfer chain, it's only a small piece. The energy bottleneck is real. Everyone's talking about it. And it's relatively straightforward to map out the companies working to solve it. By contrast, the telco network is another bottleneck that doesn't get discussed as much. Like the power transmission backbone across states, the network backbone is also sprawling and requires massive investment to upgrade network infra built during the Dotcom bubble(!) in the 1990s. But compared with power usage, where you can find public data and estimate how industries like IPPs will monetize their CapEx and energy demands, the telco network isn't as straightforward. When you send a query to ChatGPT, it travels from your device to a modem/router, then through an access network (copper, fiber, or wireless) to your ISP. For overseas traffic, it crosses subsea cables via cable landing stations (CLS), where it moves from undersea networks to terrestrial infrastructure. From there, it passes through special network hubs that include Points of Presence (PoPs) and Points of Interconnection (PoIs), where different networks exchange data to manage traffic loads. After that, the query is carried across backbone networks like long-haul/metro networks until it reaches the data center, where ChatGPT's servers process the request. For example, from Korea to the US, it typically takes about 100 ms for the query to travel one-way through these network steps to ChatGPT's servers - an engineering marvel! Of course, this is oversimplified. There are lots of technical tricks to control the data traffic, and at each node there are bottlenecks that must be optimized for latency and throughput.
- Orchestration to ensure enterprises and governments can easily adopt AI tools with maximum ROI. AI use cases are popping up everywhere. New kinds of data centers with different business models beyond the usual big tech setups are on the rise. Purpose-built ASICs particularly designed for AI inferencing to replace general-purpose NVIDIA GPUs are showing up and are just starting to be used for scaling at the enterprise level. Data sovereignty is a real deal. Security and fault tolerance are becoming must-haves. All this decentralization, along with highly complex hardware and software designs and what customers want, creates a big challenge: "how do you coordinate all the hardware and software layers to design and deploy bespoke AI clusters at scale?" Capacity isn't the only bottleneck anymore. Orchestration is starting to become a bottleneck too. Failing to manage it properly can be fatal to enterprises. The NVIDIA GB200 superchip costs around $70,000, and a sever rack including multiple chips runs about $2 to 3 million. Many such racks are installed in data centers. Imagine, after all the investment, the compute nodes just sit idle because there're no proper software/interconnect layers to coordinate across fragmented hardware to run AI workloads efficiently. Not many companies can pull this off (insanely hard!). But those that can are quietly teaming up with enterprises and data center operators.
- Water solution providers with stronger pricing power than traditional regulated water utilities, kind of like how independent power producers (IPP) operate. Nuclear, coal, and natural gas power plants use lots of water to generate energy, so as data centers and O&G companies. Natural gas is still the most realistic option to fix power demand/supply imbalances. Produced water recycling and handling is a critical, non-partisan issue that remains underappreciated by most observers. West Texas, where the Permian Basin is located, has plenty of land, (recyclable) water, and power, with little concern about community pushback. I'm curious to see if the value of the area will appreciate over the next few years among data center giants.
That said, I'm curious how the tariffs on steel, aluminum, and copper products will impact market dynamics. These three materials are the bread and butter of AI infrastructure. Transformers, which are made from many of these elements, are in severe shortage right now due to the rapid growth of data centers. About 80% of transformers are imported, and they are also subject to tariffs. Because these tariffs are imposed under Section 232 for national security reasons, it's tough for courts to overturn them. The steel and aluminum tariffs that first came in back in 2018 during President Trump's admin have stuck around through his first term and haven't been lifted by the Biden admin. But things were quite different then. U.S. inflation was low, and borrowing costs for businesses were much cheaper. Now, in 2025, we're looking at a totally different economic landscape.
According to recent earning calls by leading communication companies, after a downcycle in 2023 and 2024, wireless service providers are ramping up investment to renovate infrastructure, but rural areas may be put a a disadvantage. The Biden admin's BEAM (Broadband Equity, Access, and Deployment) program strongly favored a 'fiber-first' approach for accessing high-speed internet, but the policy has been significantly updated to replace the fiber-first stance with a 'technology-neutral' approach like satellite communications. Installing fiber optics networks is more expensive in rural areas than in urban centers, and rising component prices will further increase costs.
That being said, 5G networks are being pursued as effective rural broadband solutions. The One Big Beautiful Bill (OBBB) Act will speed up 5G and private network rollouts. The Trump admin restored the FCC's authority to auction spectrum, which had expired in 2023, and reinstated 100% bonus depreciation, allowing MNOs and enterprises to write off the full cost of 5G network equipment in the year of purchase. Starting in 2026, FCC will auction substantial amounts of mid-band and low-band spectrum over the next 8 years, which is key to enabling a nationwide 5G network. This is a huge catalyst for 5G network equipment providers.
Besides AI infrastructure, I've been closely following developments in quantum sensing, and more broadly, in Positioning, Navigation, and Timing (PNT), especially as 5G/6G frequency bands overlap or sit right next to military bands like the lower three GHz band used for missile defense. Quantum sensing can complement or serve an alternative to traditional GPS. Unlike GPS, it is anti-jamming, anti-spoofing, and capable of detecting extremely weak signals. Most of the buzz is around quantum computing and startups that lack serious technology. PNT doesn't sound as exciting as AI and robotics, but GPS (an "invisible utility") underpins ground/aviation/maritime transport, financial transactions, power grid operations, agriculture, LEO satellite synchronization, and even AI and robotics. Disruption is already real. 85% of flights in Estonia experience GPS disruption because of Russia's electronic warfare against Ukraine. It is even affecting LEO satellites as they pass over Ukraine. In late 2024, the DoD raised the budget limit to use commercial LEO constellations from $900 million to $13 billion. I'm sure the DoD won't be pleased if the same kind of disruption starts hitting U.S. assets.
The Achilles' heel of drone warfare is the lack of signal resilience against EMI. PNT is now slowly coming to the forefront of investors' attention for enabling two key initiatives: Golden Dome and LEO satellite proliferation. Detecting hypersonic missiles and managing satellite and even emerging eVTOLs traffic requires high-precision time and frequency measurement on the picosecond scale, or less than a billionth of a second. Quantum sensors are (quietly) being field-tested in the defense and space sectors. 2025 marks an inflection point, with the focus now on ruggedized quantum sensors agnostic to military platform and environment beyond the static lab setting. It's exciting to see how quickly the space ecosystem is evolving.
Speaking of LEO satellite proliferation, China is on track to complete the first phase of the SpaceSail Project by the end of 2025, with a target of launching 15,000 LEO satellites by 2030 as a direct competitor to SpaceX's Starlink. The rivalry is no longer limited to AI and semiconductors. Space has become a new arena in the escalating U.S.-China competition. The space industry is characterized by a 'first come, first served' dynamic and is key to the global digital order - no matter what the ultimate goal may be (which is kind of scary). Personally, I enjoy analyzing the space industry's value chain, from satellite manufacturers to operators, ground equipment suppliers, service providers (MNOs), and end users. It's quite vibrant; for example, advanced terminals for in-flight connectivity (IFC) are line-fitted or retrofitted on lots of commercial and business aircraft annually.
This is my own half-baked philosophy, but I think that the common thread running through themes like AI, robotics, advanced computing, communications, missile defense systems, space technology, and quantum science is PNT. It's actually the biggest single point of failure in all of these fields. In the era of software-defined everything (SDx/SDE) across the commercial, civil, and defense sectors, a tiny misstep like a split second error or PNT interference can have a huge impact on machine efficiency and human lives.
Whoever productizes hardware and software that transport electrons and photons faster, denser, with less loss, lower power, lower noise, tighter sync will ride this AI era. A perfect example is Co-Packaged Optics (CPO), now integrated into the Blackwell platform to replace power-hungry Digital Signal Processor (DSP). On top of that, Nvidia is reportedly exploring a new packaging architecture called CoWoP, set to replace the current CoWoS used across all GPUs, including Blackwell. This approach removes the substrate (S), mounting the chip interposer (CoW) directly onto the PCB (P). Easier said than done. It puts immense pressure on PCB value chains. Resonac, a key player, has already formed a consortium with leading semi companies to address it.
Lord Kelvin once said, "If you can't measure it, you can't improve it." Look at the LLM benchmarking wars, where leaders clash and leapfrog each other almost weekly or monthly. The ranks basically reflect who has the most advanced system to measure and improve the flow of electrons and photons, the fundamental carriers of data.
In the age of AI inferencing, I think the real frontier may be the network - the layer that governs how electrons and photons flow at every scale - server, rack, cluster, data center fabric (Broadcom, Arista, Nvidia, Credo, Coherent, Marvell, Lumentum, MaxLinear), inter data center (Ciena), national backbone (AT&T, Verizon, Equinix, Lumen), intercontinental, space (SpaceX, Amazon, AST SpaceMobile, Viasat), and ultimately edge (Akamai, Cloudflare). From backhaul to fronthaul, from first mile to last mile, numerous competitors are trying to eat each other's lunch within this fragmented network market.
Unless you're interested in technology purely for its own sake, staying informed about macroeconomic trends and geopolitics is essential to understanding how the technology landscape evolves. Investment volumes shift dynamically, and a vertical that was once second-tier in the global market can become first-tier, and the reverse can also happen.
For example, while jamming and spoofing pose real threats to commercial airlines, the development and deployment of quantum sensors depend largely on federal funding for defense applications. Monitoring the defense budget and understanding how funds are allocated across different focus areas is critical. I enjoy reading abstracts of awardees in the Small Business Innovation Research (SBIR) program to learn about early-stage R&D activities by small businesses. The DoD is an attractive target market, as it adopts new technologies quickly and is flexible about procedures or company size as long as performance is validated.
In FY 2026, sensing capabilities and satellite communications are the top priorities in support of the Golden Dome initiative. Photonics and PNT technologies, which I mentioned earlier, are central to these efforts. Also, space-grade alloys and composites are needed to intercept missiles and reduce payload weight. Just as critical is the systems engineering that ties everything together. It's like owning a stack of high-end Nvidia GPUs makes no sense if you don't have the right abstraction layer to direct and route the flow of data. It's quite fun to learn all these advances (though it is kind of scary).
Lastly, a lot of folks are talking about market bubbles, potential collapses, and recessions. The current trend seems to be that whenever (minor) corrections happen, massive dip-buying follows. Many investors don't seem scared of corrections anymore. They see them as chances to scoop up bargains. The U.S. has been on an easy money steroid shot since 2008, driven by multiple rounds of quantitative easing (QE), years of ultra-low interest rates, and lavish fiscal policy. Even after 3 years of quantitative tightening (QT) since mid-2022, there's still plenty of money (M2) circulating around, which has kept the market artificially propped up. The Fed's balance sheet shows a correlation with the S&P500.
Are market valuations stretched? Yes. Will the U.S. economy hit a recession? Recent macro data aren't pointing in that direction yet. BUT, major corrections triggered by disruptive policy changes and geopolitical events could cause a recession, rather than by themselves. That's just my armchair take, with all the risk of being totally wrong and facing the humiliation! I'm not a trained economist or a professional market analyst, just someone who thinks it's worth building a rough framework even as a layperson.
Here are some useful links to help follow macroeconomic trends for laypeople like me.
- econ p.i. - provides "big picture" insights into economic conditions by aggregating key economic data and sentiments.
- Wolf Steet - offers commentary on economic indicators and broader macroeconomic trends.
- Dodge Momentum Index - a 12-month leading indicator for non-residential construction spending.
- The Association for Manufacturing Technology - reports new orders of metalworking machinery with about a 2-month delay.
- OpenTable - tracks seated diners change year-over-year, offering sentiment insights different from MSCI.
My framework for understanding market trends is built on irreversible, inevitable forces, independent of contentious AI bubbles.
- The era of abundance and globalization is stalling, burdened by devaluing currencies, aging populations, and ballooning debt. Tolérance is exhausted. Nations are radicalizing. A zero-sum power struggle is underway. The U.S. is restructuring the China-centered global supply chain. Tariffs are a byproduct. Even if the tariff card wasn't played, other measures would've come out. Great powers are doubling down on technological advancement to boost productivity and reshape the global order.
- Compute and energy are becoming sovereign assets, strengthened by the electro-industrial stack (EV, robotics, automation).
- Space is becoming both a warfighting domain and a commoditized marketplace.