David Shapiro · Tech & AI
TIER 4 Thu, 29 Jan 2026 14:12:29 +0000
Watch now (26 mins) | The expected (an unexpected) friction that is bogging down AI progress ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ ͏ | | ---|---|--- | | | Forwarded this email? Subscribe here for more --- --- --- | | Watch now --- # Why AI is slowing down in 2026 ### The expected (an unexpected) friction that is bogging down AI progress | | David Shapiro --- | Jan 29 --- | --- --- | | | --- | | --- | | --- | | --- | | READ IN APP --- As we enter 2026, a curious gap has emerged between the breathless predictions of artificial intelligence enthusiasts and the reality unfolding on the ground. The technology itself continues advancing at a remarkable pace. The research papers keep coming. The models keep improving. Yet something is holding back the full realization of AI's potential, and the reasons have almost nothing to do with the debates dominating public discourse. The bottlenecks are physical, structural, and in some cases almost absurdly mundane. Understanding where the friction actually lies reveals a very different picture of the AI landscape than what most commentators present. The short version of the story goes like this. The most significant barriers to AI acceleration are thermodynamic in nature, meaning they relate to power generation, grid capacity, and the physics of getting electricity where it needs to go. Next comes the supply chain, particularly the packaging and memory components that complete the hardware stack. Further down the hierarchy sits operational friction, encompassing the challenges enterprises face in actually deploying AI systems, managing data quality, and demonstrating return on investment. And at the very bottom, contributing remarkably little actual friction, are the concerns that dominate headlines and social media debates. Safety discussions, ethical hand-wringing, and regulatory posturing turn out to be largely noise when measured against the real constraints preventing AI from moving faster. This is a frustrating reality for those who have spent years engaged in philosophical debates about AI alignment and existential risk. The actual conversation sphere has not caught up with the operational reality. Serious researchers and policy people with actual decision-making power have largely moved on from the abstract safety debates that once commanded so much attention. The concerns have been operationalized into standard risk management frameworks, and the companies building AI are accelerating regardless of what any particular commentator might say on social media. ## The Magnitude of Investment and Why It Still Falls Short Before examining the bottlenecks themselves, it helps to appreciate the sheer scale of resources flowing into artificial intelligence. The assumption driving much of this investment has been elegantly simple. If we need something, we can buy it. Money, in sufficient quantities, should dissolve any obstacle. This premise has unleashed a flood of capital unprecedented in the technology sector. Wall Street consensus estimates for hyperscaler capital expenditure in 2026 now stand at approximately 527 billion dollars, up from 465 billion dollars at the start of the recent earnings season. This continues a trend of upward revisions that has persisted for years. Major technology companies including Amazon, Alphabet, Microsoft, and Meta are collectively on track to invest over 400 billion dollars into AI initiatives in 2025 alone, and they have uniformly indicated that this spending is nowhere near enough. Amazon has announced capital expenditures of 125 billion dollars for 2026, with the vast majority dedicated to AI and related infrastructure for Amazon Web Services. Microsoft disclosed roughly 80 billion dollars for its fiscal 2025, with more than half destined for facilities within the United States. Alphabet signaled 75 billion dollars as demand mounted through the year. Meta's capital expenditures have nearly doubled from the previous year to approximately 72 billion dollars, with executives indicating the 2026 figure will grow notably larger still. Gartner projects total global AI spending will reach nearly 1.5 trillion dollars in 2025 and exceed 2 trillion dollars by 2026. These figures include software, devices, and infrastructure across all segments of the market. Stanford's AI Index reported that corporate AI investment reached 252.3 billion dollars in 2024, with private investment climbing 44.5 percent and mergers and acquisitions up 12.1 percent from the previous year. The sector has experienced dramatic expansion over the past decade, with total investment growing more than thirteenfold since 2014. US private AI investment hit 109.1 billion dollars in 2024, nearly twelve times higher than China's 9.3 billion and twenty-four times the United Kingdom's 4.5 billion. When you add it all together, AI investment now rivals the economic commitment of the Manhattan Project as a share of GDP. We are living through what amounts to a decentralized version of those historic mobilizations that defined the twentieth century. The interstate highway system. The Apollo program. The original Manhattan Project. The current AI buildout matches them in relative scale and ambition. Yet all that money keeps running into walls that capital alone cannot breach. The economy is restructuring around these new problems, but the problems themselves remain stubbornly physical. You cannot purchase your way past the laws of thermodynamics, the lead times on specialized equipment, or the fundamental constraints of electrical grid infrastructure built over decades for a very different world. ## The Energy Crisis Nobody Is Talking About Energy represents the hardest stop in the entire AI acceleration stack. The numbers here are staggering enough to seem almost unbelievable, yet they emerge from straightforward analysis of current projects and planned expansions. The wait time for connecting a new data center to the American electrical grid now commonly stretches to five to seven years for large facilities. Developers in Northern Virginia, the country's largest data center hub, report seven-year delays while lawmakers explore stricter siting rules and shared upgrade costs. That timeline would place new facilities coming online in 2032 or 2033, by which point the most optimistic roadmaps suggest we should have achieved forms of artificial general intelligence. The disconnect between technological ambition and infrastructure reality could hardly be more stark. To understand the scale of demand, consider that the United States had approximately 25 gigawatts of operating data center capacity in 2024. According to analysis from S&P Global and 451 Research, US data center demand will rise to 75.8 gigawatts in 2026 for IT equipment, cooling, lighting, and other uses. That figure expands to 108 gigawatts in 2028 and 134.4 gigawatts by 2030. BloombergNEF released even more aggressive estimates in late 2025, projecting US data center power demand could reach 106 gigawatts by 2035, a forecast 36 percent higher than their previous prediction from just months earlier. For perspective, a typical nuclear reactor produces about one gigawatt of power. The AI industry is trying to conjure the equivalent of more than 130 nuclear plants worth of power capacity in roughly six years. These figures seemed implausible at first glance. Surely the projections must be inflated by hype or wishful thinking. But when you examine the actual projects in permitting queues and the backlog of orders already placed, the numbers hold up. The Stargate project alone, announced by OpenAI with backing from Microsoft and other partners, involves infrastructure and technology investments approaching one trillion dollars. The first gigawatt of capacity from their Nvidia and AMD deals is expected to begin deployment in the second half of 2026, with plans for 16 gigawatts from those deals plus another 10 gigawatts of custom-designed accelerators from Broadcom. This represents not hypothetical future demand but rather what companies are actively trying to build right now. The solutions emerging from this constraint involve bypassing the grid entirely where possible. Microgrids combining solar panels, battery storage, and natural gas turbines on site can provide power without waiting for interconnection permits. Some companies are exploring direct connections between nuclear plants and data centers, removing the grid from the equation altogether. Microsoft announced a 1.6 billion dollar deal to revive the Three Mile Island nuclear reactor specifically to power AI infrastructure. Google and Amazon have secured similar nuclear energy agreements. The pattern emerging across the industry involves major technology companies transforming themselves into de facto utility providers, generating and managing their own power rather than relying on existing infrastructure. Historical precedents exist for this kind of arrangement. In Russia's far east, aluminum smelting operations run on dedicated hydroelectric facilities because the energy costs of smelting make proximity to cheap power the dominant factor in plant location. America lacks Siberia's vast empty expanses, but it does have enormous deserts well suited for solar generation at scales that could feed dedicated AI facilities. ## The Physics of Waiting The grid interconnection problem traces back to physical equipment with extraordinarily long lead times. Large power transformers now average lead times of approximately 128 weeks, or about two and a half years. Generator step-up transformers run even longer at roughly 144 weeks, approaching three years. Some facilities have reported wait times ranging from 80 to 210 weeks depending on specifications, with one large power transformer manufacturing facility disclosing a five-year wait for new orders. Wood Mackenzie's analysis projects that the United States will face a 30 percent shortfall in power transformers and a 10 percent gap in distribution units through 2025 and into 2026. Demand for high-voltage power transformers and generator step-up transformers has surged by 116 percent and 274 percent respectively since 2019, far outpacing what domestic manufacturing capacity can handle. Substation demand increased by another 91 percent during that period. The situation stems from decades of underinvestment in domestic manufacturing combined with a sudden surge in post-pandemic construction and electrification. Volatility in grain-oriented electrical steel and copper markets has pushed lead times for large power transformers well beyond historical norms. The United States relies on a single domestic supplier of the specialized electrical steel used in transformer cores, forcing manufacturers to source raw materials or semi-finished components from abroad. These are not the small transformers visible at neighborhood substations. These are house-sized units that most people never see, weighing between 100 and 400 tons and requiring specialized transport. Sometimes delivery requires one of approximately ten suitable super-heavy-load railcars in the entire country. Those logistics alone can add months to a replacement project. This creates a crisis window running from approximately 2026 through 2028. During this period, the infrastructure projects and equipment orders now being placed will slowly work through the system, but the gap between demand and supply will remain acute. Nuclear power, often discussed as the ultimate solution to AI's energy needs, offers no relief during this window. Even the most optimistic timelines for restarting dormant plants or constructing new ones push meaningful delivery into the 2030s. Small modular reactors, the darlings of nuclear optimists, will arrive too late to address the immediate crunch. What can spin up quickly enough to matter? Natural gas turbines, solar installations, and grid-scale batteries represent the realistic near-term options. Iron-air battery technology has emerged as a particularly promising solution for grid-scale storage. These batteries sacrifice efficiency and portability for dramatic cost reductions and exceptional longevity. When your battery can be as heavy as needed and sits permanently in one location, the calculus changes entirely. The common objection that solar cannot cover baseload demand misses this point. With sufficient battery capacity, the intermittency of solar generation becomes manageable. The question shifts from whether solar can work to whether batteries can be manufactured and deployed at the necessary scale. ## The Supply Chain Has Moved On Two years ago, the technology press obsessed over GPU shortages. Nvidia chips became objects of almost mythical scarcity, with wait times stretching months into the future and secondary market prices reaching absurd heights. That particular bottleneck has largely resolved itself through the normal operations of markets responding to price signals. The shortage has migrated elsewhere in the stack. High bandwidth memory now represents the critical constraint. All three major suppliers have essentially sold out through 2026. SK Hynix reported during its October 2025 earnings call that its HBM, DRAM, and NAND capacity is sold out for 2026. Micron announced during its fiscal first quarter 2026 earnings that its high-bandwidth memory capacity is sold out through calendar year 2026. Samsung faces similar constraints despite being the world's largest memory manufacturer. The market has responded with unprecedented price increases. TrendForce reported that average DRAM memory prices are expected to rise between 50 and 55 percent quarter over quarter, a jump that analysts described as unprecedented in the industry's history. Samsung raised prices for 32-gigabyte DDR5 modules from 149 dollars to 239 dollars, a 60 percent increase. Contract pricing for DDR5 has surged more than 100 percent. The companies capable of manufacturing this memory have rationally shifted production away from consumer applications toward AI accelerators, where margins run considerably higher. HBM requires significantly more wafer capacity per bit than standard DRAM modules, meaning every wafer pushed into HBM production removes capacity from conventional memory for PCs, smartphones, and consumer electronics. OpenAI and Microsoft finalized preliminary agreements with Samsung and SK Hynix for chip supplies for the Stargate project, with the memory manufacturers committing to up to 900,000 DRAM wafer starts per month to support the buildout at an accelerated capacity. Technology companies including Google, Amazon, Microsoft, and Meta Platforms have placed open-ended orders with memory suppliers, indicating they will accept as much supply as available regardless of cost. The downstream effects are rippling through consumer electronics. Chinese firms including Xiaomi have warned of impending price increases for mobile devices. In Tokyo's Akihabara electronics district, retailers began limiting purchases of memory products to prevent hoarding. Nvidia announced plans to slash RTX 50-series gaming GPU production by 30 to 40 percent in the first half of 2026 due to GDDR7 shortages. Micron has announced it will exit the Crucial consumer memory brand entirely to focus on enterprise and AI customers. Executives stated that winding down the consumer business would free up wafer supply for strategic accounts. The memory shortage is expected to persist through at least late 2027, with new fabrication facilities not becoming operational until 2027 through 2030. Beyond memory sits the packaging constraint. The complete chip-on-wafer-on-substrate assembly process has become another limiting factor in delivering finished AI accelerators. Nvidia alone now books more than half of global packaging capacity for this advanced assembly work, leaving competitors scrambling for the remainder. SK Hynix has told investors that its advanced packaging lines are at capacity through 2026. The GPU die itself, the logic component that once seemed like the obvious chokepoint, no longer limits production. Market dynamics will eventually resolve these supply chain issues just as they resolved the earlier GPU shortage. The timeline looks similar as well. Expect 18 to 24 months before the memory and packaging constraints begin to ease meaningfully. The market will sort this out. It just takes time. ## The Data Question People keep talking about data as a constraint on AI development. The argument holds that the internet has been scraped, the books have been digitized, and we are running out of raw human-generated content to feed into training pipelines. This concern deserves mention but may be overstated. The latest generation of AI models are trained on increasing proportions of synthetic data. While AI trained exclusively on its own outputs does tend to result in model collapse, researchers have developed increasingly sophisticated techniques for generating and validating synthetic training data that maintains quality. More fundamentally, humans do far more with far less data. A human brain develops general intelligence from perhaps twenty or thirty years of experience that represents less than one percent of the data volume used to train current large language models. We are clearly doing something different. We are not getting the most out of the data that we have. If necessity is the mother of invention, constraints are the father of creativity. Should we genuinely run out of novel data, the research community will find better algorithms to extract more learning from existing sources. We clearly have enough data to generate human-level performance on many tasks. The algorithms simply are not yet efficient enough to learn as effectively as biological neural networks. ## The Six Hundred Billion Dollar Question Here sits the most interesting tension in the current AI landscape. Investors are growing wary of writing hundred-billion-dollar checks for data center infrastructure when returns on investment remain years away. The hyperscalers that raised massive debt in 2025 are watching investors look at them with increasing skepticism. This explains why executives like Sam Altman have turned to sovereign wealth funds in places like Saudi Arabia seeking financing for data center expansion. The average stock price correlation across the large public AI hyperscalers has declined from 80 percent to just 20 percent since mid-2024. Investors have rotated away from AI infrastructure companies where operating earnings growth is under pressure and where capital expenditure is being funded via debt. At the same time, investors have rewarded companies demonstrating a clear link between capital expenditure and revenues. Bank of America credit strategists noted that the five largest megacap technology companies collectively may be reaching a limit to how much AI capital expenditure they are willing to fund purely from cash flows. Consensus estimates suggest AI capital expenditure will climb to 94 percent of operating cash flows, minus dividends and share repurchases, in 2025 and 2026. That figure is up from 76 percent in 2024. They do not need to borrow to fund spending yet, but the margin for error is shrinking. The question hangs in the air. Can spending at this scale sustain belief alone? We are in a period resembling Solow's paradox or the J-curve of productivity, where AI is proliferating through the economy but the productivity gains have not yet shown up clearly in the statistics. Most companies are using AI now. In 2024, the proportion of survey respondents reporting AI use by their organizations jumped to 78 percent from 55 percent in 2023. But the technology is not saturated to the degree investors would prefer. This mirrors the early personal computer era of the 1980s, when businesses had machines on their desks that could connect to central databases and perform useful tasks but had not yet achieved anything close to their eventual transformative potential. It also mirrors the early internet, when digital subscriber lines and dial-up connections made the network useful but expensive and limited. The cost per token remains too high for many applications. The maturity of the tools remains insufficient for many use cases. Enterprise return on investment is not manifesting as quickly as executives would like. Studies consistently find that the vast majority of AI pilot projects fail to reach production. Data quality problems trip up implementation efforts. Integration with legacy systems running on operating systems that are twenty or thirty years old proves far more difficult than anticipated. The lack of qualified AI engineering talent creates bottlenecks that money alone cannot solve. One estimate suggests there are globally only about 22,000 high-end AI engineers, which is simply not enough to staff the buildout the industry envisions. The barriers are very mundane. Cost. Return on investment. Data governance. Technical debt. None of this has anything to do with safety or ethics debates. ## The Surprising Role of Insurance Here is perhaps the most ironic finding to emerge from analysis of AI deployment friction. Many insurance policies now include absolute AI exclusions. This means that if a company uses AI and there is an OSHA violation or a patient dies or any number of adverse outcomes, the insurance company completely washes its hands of the incident. If AI touched it, the insurer will not take any responsibility whatsoever. The reason is straightforward. AI is new. It is high risk. It is high variance. Insurance companies do not know how to price it. When you have a domain that is well understood and the risks can be controlled and measured, underwriters know how to create policy templates that can be sold at scale. They have actuarial tables and historical data and established frameworks for managing liability. None of that infrastructure exists yet for AI-related risks. Creating bespoke policies for novel risk categories is possible but extraordinarily labor-intensive. It cannot be operationalized or scaled in the way that standard insurance products can. So instead of figuring out how to price AI risk, many insurers have simply excluded it from coverage entirely. The downstream effect is that enterprises that would otherwise adopt AI solutions decline to do so because they cannot obtain adequate insurance coverage. No insurance means exposure to potentially catastrophic liability, which makes risk-averse organizations unwilling to proceed regardless of the potential benefits. This may be the dumbest reason to slow down AI adoption. The technology works. The business case is compelling. The competitive pressure is intense. But legal departments say no because the insurance markets have not figured out how to write policies for it yet. If any insurance professionals are looking for an opportunity to make a meaningful contribution to AI acceleration, this represents a genuine gap that someone needs to fill. ## The Regulatory Reality Most public conversation about AI regulation focuses on safety requirements, consumer protections, copyright issues, and concerns about AI-generated content. These debates consume enormous amounts of attention and emotional energy. They matter almost not at all to the actual pace of AI development. Frontier labs continue accelerating despite high-profile researcher resignations. Why? Because there is a lot of money to be made if you are a talented AI engineer, and that will remain true for quite some time. Business-to-business adoption proceeds without any regard for public sentiment about AI art or concerns about copyright infringement. Enterprises do not care about those debates. The concerns do not impact their decisions whatsoever. US federal regulation remains largely noise. The regulatory agencies have neither the technical expertise nor the political will to meaningfully constrain AI development. What little attention they pay to the sector tends to focus on consumer protection issues that lie far downstream of the core technology development. Europe is a different story. The EU AI Act functions as a deliberate brake on development, creating a compliance wall that effectively excludes startups from high-risk AI use cases. Licensing costs for any high-risk systems run around 52,000 euros annually, which means no early-stage company will operate in that space within European borders. The result is predictable. Those companies go elsewhere. They go to America. Some go to China. Some go to Saudi Arabia. Europe has become exceptionally good at ensuring that frontier technology development does not happen within its borders. The regulatory capture and vetocracy that characterize European institutions mean that the continent will watch the AI revolution unfold from the sidelines. Where regulation does create genuine friction is in the geopolitical competition between America and China. Export controls and import controls represent the real regulatory barrier at the global level. America already has perhaps two to three to five times the compute capacity that China possesses. That gap is widening. By 2027, America is expected to have seventeen times the compute of China, a disparity driven largely by export controls on advanced semiconductor equipment. From a geopolitical perspective, this means America is winning. That compute advantage represents the nation's moat. Chinese models remain inferior across the board according to assessments from the Center for AI Security and other research organizations. Some Chinese systems are cheaper or more efficient, but in terms of raw capability and security risk profiles, they do not represent a competitive threat at the military or strategic level. Of course, constraints are the father of creativity, and the Chinese have proven remarkably capable of doing more with less. But the structural advantages America enjoys will take years for competitors to overcome, if they can overcome them at all. ## The Digestion Phase What we are living through has been called the digestion phase of AI development. The hype cycle ran from 2023 through 2025. Now reality is catching up. Reality says we need greater interconnection capacity. We need transformers and substations and transmission lines. We need high bandwidth memory and advanced chip packaging. We need verifiable synthetic data and better training algorithms. We need insurance products that can actually cover AI-related risks. We need engineers who know how to build and deploy these systems at scale. The paradigm has shifted from bigger models and scale is all you need to efficiency and distillation and making do with what exists. Do the best with what you have. That is the operating philosophy for the next couple of years. This is not to say that no new capacity is coming online. Obviously capacity is expanding constantly. But the distance between what the industry would like to install and deploy versus what can actually be deployed continues to grow. The demand curve is outpacing the supply curve, and the gap will persist until approximately 2028 when infrastructure investments currently being made begin to mature. After 2028, acceleration can resume in earnest. The economic pivot will continue. The friction that currently constrains development will ease as markets work through the backlogs and infrastructure catches up with ambition. ## A Map of What Actually Matters To summarize the friction landscape as it exists today. The critical and binding constraints are power availability, grid interconnection, and high bandwidth memory supply. These represent hard physical limits that cannot be wished away or purchased through. They require time and infrastructure investment that operates on multi-year timescales. The moderate and addressable constraints include data quality challenges that can be solved through better engineering practices, packaging capacity limitations that the market is actively working to resolve, deployment friction that diminishes as enterprises learn to operationalize AI effectively, and liability and insurance gaps that require new products and frameworks but face no fundamental obstacles. The things that are not actually constraints at all include safety and existential risk discussions, federal regulation, and capital availability. Money is abundant. Regulatory pressure is minimal. The philosophical debates about AI alignment have been operationalized and incorporated into standard risk management. None of these slow down development in any meaningful way. This represents a very different picture than the one presented in most public discourse about AI. There is no bubble to speak of, at least not in the traditional sense. Everything is sold out. Demand exceeds supply across nearly every relevant category. With a traditional speculative bubble, you have pure speculation driving valuations beyond any reasonable assessment of fundamental value. With AI, people want more than the industry can give them, and the technology is not even close to fully matured yet. The future belongs to those who master the physical world. Grid permits. Fabrication capacity. Energy generation. Transformer manufacturing. Memory production. The research frontier continues advancing at pace. The science is not the bottleneck. Now we have entered the phase where rubber meets road, where the friction of physical reality becomes the binding constraint. Stop arguing about philosophy and start pouring concrete. You're currently a free subscriber to David Shapiro's Substack. For the full experience, upgrade your subscription. Upgrade to paid --- | | | Like --- | | Comment --- | | Restack --- (C) 2026 David Shapiro 548 Market Street PMB 72296, San Francisco, CA 94104 Unsubscribe