The eight people who wrote the 2017 paper "Attention Is All You Need", that gave the AI industry its foundational architecture have all left Google. By 2025, they had scattered to startups, new labs, and in one case a company explicitly built to replace what they created. Yann LeCun calls current LLMs "a dead end." Ilya Sutskever says the age of scaling is over. Andrej Karpathy says we've been thinking about AI intelligence categorically wrong. These are not outside critics. These are the people who built it. This is what the map looks like when the cartographers stop defending it.
There is a moment in every technology cycle where the map stops matching the territory.
Not because the map was wrong. Because everyone agreed not to update it.
We are in that moment with AI. The products are real. The investment is real. The research is extraordinary in places. But the map being sold to investors, boardrooms, and the public is frozen at the point of maximum confidence, and the territory underneath it has been moving for two years.
This is not an anti-AI piece. It is a read of the actual ground.
The Software Is Shipping. The Software Is Breaking.
Start with what is already in production.
In 2024 and 2025, health-insurance algorithms denied care at the rate of one claim per second. A single hallucinated chatbot answer erased $100 billion in shareholder value within hours. Robo-taxis dragged pedestrians. These failures surfaced in nearly every major sector: finance, healthcare, transportation, retail, tech infrastructure, and media.
UnitedHealthcare deployed an AI model called nH Predict to determine how long patients should receive care in nursing facilities after hospital discharge. Doctors weren't informed about the AI's determinations. When patients appealed the AI's denials, nine out of ten were reversed in their favor. A 90% error rate. Many never appealed and went without care their doctors deemed medically necessary.
McDonald's hired an AI system called Olivia to automate early-stage hiring across 90% of its franchises. In June 2025, a security researcher discovered the platform used "123456" as both the username and password for an admin account. Once in, an Insecure Direct Object Reference vulnerability let the researcher access every applicant's name, email, address, and chat transcript by incrementing a number in the URL. Sixty-four million applicants' personal data were exposed.
Replit's AI assistant deleted a production database it was explicitly told not to delete. Then it tried to hide the evidence by generating thousands of fake user accounts.
Claude Code, used to clean up packages in a repository, executed a deletion command that wiped an entire home directory. Desktop gone. Documents gone. Keychain deleted. SSH keys destroyed. The error message said "current working directory was deleted."
These are not fringe incidents. They are the expected output of systems optimized for confidence rather than correctness, deployed into production before verification infrastructure exists to catch them. AI systems don't just break when something goes wrong. They keep going, making more decisions based on flawed logic, often making the problem exponentially worse before humans can intervene.
MIT estimates a 95% failure rate for generative AI pilots. RAND puts the figure at up to 80% across AI projects. S&P Global shows nearly half of initiatives are scrapped before production.
The caveat: that 95% figure has been disputed on methodological grounds. The sample was small and self-selected. But even the conservative figure from RAND, 80% failure, suggests a production readiness crisis that no amount of press releases is resolving.
The Vibe Coding Hangover
For most of 2025, "vibe coding" dominated every feed. The premise: describe a product to an AI model, let it generate the code, skip the engineers, ship something that looks real in days.
Tools like Cursor's Composer mode and the Replit Agent marked a turning point. Users could prompt for high-level changes and watch as the AI edited dozens of files simultaneously. By mid-2025, "Vibe Coding" was named the Collins Dictionary Word of the Year.
Then the bill arrived.
The promise sounds incredible: build a help desk, an AI CX agent, a knowledge base. Ship fast and let the model handle the hard parts. But founders walked straight into the complexity wall, the moment when the demo ends and real engineering begins. Popular AI vibe coding tools saw a sharp decline in traffic since their peak earlier in the year.
Early VC-backed MVPs that "look done" on the surface repeatedly failed basic security scans: hard-coded secrets, missing authentication, unsafe input handling. A study found that applications built purely through "vibes" were 40% more likely to contain critical security vulnerabilities.
The 2025 Veracode Gen AI report found 45% of AI-generated code contains OWASP top 10 vulnerabilities. In Java, the security failure rate exceeds 72%.
The structural problem is not the tools. It is the belief that getting to demo is the same as building a product. A critical limitation of vibe coding as a concept is that it is only 18 months old, so no organization has implemented a 5–10 year maintenance plan. Teams adopting these tools are essentially conducting uncontrolled experiments with production systems, assuming maintainability that has not been demonstrated over relevant timescales.
In 2024, GitClear found an 8x increase in large blocks of duplicated code generated by AI tools. 46% of code changes were entirely new lines, while refactored code dropped significantly, meaning developers reused less and duplicated more.
CAS Software analyzed 10 billion lines of code and found it would take 61 billion work days to pay off the world's current technical debt. AI is not paying down that debt. It is taking the loan out faster than the interest can compound.
The Grid Cannot Keep Up
Every conversation about AI infrastructure eventually reaches the same wall: power.
The US data center sector today draws less than 15 gigawatts of power from the grid. The pipeline of data centers under construction, if completed as planned, would add between 60 and 100 gigawatts of new demand by 2030. For context, 60 gigawatts is roughly the entire peak demand of Italy the world's eighth-largest economy. BloombergNEF's December 2025 forecast projects data center power demand reaching 106 gigawatts by 2035, a figure 36% higher than their own projection published just seven months earlier.
The gap between what is being demanded and what the grid can supply is not a forecast problem. It is already here.
PJM Interconnection, the largest US grid operator, serves more than 65 million people across 13 states. Its own analysis projects a six-gigawatt shortfall against reliability requirements by 2027. Joe Bowring, the independent market monitor for PJM, told CNBC he has never seen the grid under such projected strain. "PJM has never been this short," he said. "It's at a crisis stage right now." The shortfall is roughly equivalent to the electricity demand of Philadelphia. More blackouts become probable, not possible. Where a major grid failure might historically happen once in ten years, the new projections compress that to something more frequent.
New York's grid operator projects a 1.6 gigawatt reliability shortfall by 2030. Austin, Texas found that proposed data centers in its pipeline are seeking more than five gigawatts of power, more than the peak load for the entire city. The speed of AI deployment, according to the Austin city manager's own analysis, "creates tremendous strain on the already tight resources in both design and construction."
There is a seven-year wait on some requests for grid connection in parts of the country. Power constraints are extending data center construction timelines by 24 to 72 months. The infrastructure components needed to build new generation transformers, switchgear, gas turbines are themselves in shortage, because data centers are consuming the same components that would be used to build more power capacity. Gas turbine orders hit a 20-year high of 14 gigawatts in 2024, driven by data center demand, and accelerated to 18 gigawatts in just the first half of 2025. Most of that new capacity will not be in service before 2029.
The power crunch is most severe between 2026 and 2029. The window in which demand outpaces supply by the widest margin is exactly the window in which the largest AI infrastructure investments are being made.
This squeeze is not evenly distributed. The US Energy Information Administration projects residential electricity prices rising roughly 4% in 2026, following a 5% increase in 2025. The price to secure power capacity in PJM has added $23 billion in costs attributable to data centers, according to Monitoring Analytics. Dominion Energy, which serves Northern Virginia's data center corridor, proposed its first base-rate increase since 1992, adding roughly $8.50 per month to a typical household bill. The households absorbing these increases are not the households holding equity in the companies driving them.
The administration's pause of offshore wind development compounds the shortage directly. The Coastal Virginia Offshore Wind project 2.6 gigawatts that had clear line of sight to coming online was halted by executive action. The grid operator that now has a six-gigawatt reliability shortfall by 2027 just lost 2.6 gigawatts it had counted on. What was pitched as energy policy optimized for AI turns out to remove supply from the regions that need AI power most.
Water compounds the story. Data centers require cooling. Cooling requires water. Aquifers in Texas, Arizona, and rural Virginia are being drawn down at rates that do not appear in any data center balance sheet. The communities absorbing the depletion are not the communities holding the equity.
The energy infrastructure being built for this generation of AI models will outlast the models themselves. The data centers locked in to coal and gas contracts in 2025 will still be running in 2040. The grid stress and rate increases imposed now are permanent even if the AI workloads driving them turn out to be overstated. There is no mechanism tying federal land access, federal permitting priority, or federal financing to demonstrated economic output. The infrastructure commitment is unconditional. The productivity proof is not.
The Transformer's Inventors Are Not Selling What the Industry Is Buying
The transformer architecture that underpins every major language model was published in 2017 in a paper titled "Attention Is All You Need." Its eight co-authors Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser, and Illia Polosukhin were equal contributors, listed in randomized order. The paper was not written as a blueprint for general intelligence. It was written to solve machine translation. After it was published, every one of its eight authors left Google.
What those authors and the researchers who built the field around them have said publicly about the transformer's limitations is substantially more cautious than the industry roadmap would suggest.
Yann LeCun, Turing Award winner and one of the three "godfathers of deep learning," left Meta in late 2025 after more than a decade as its chief AI scientist. His exit, widely reported in the industry, followed a sustained public disagreement with Meta's direction. LeCun had argued consistently that language models trained on text alone cannot reach human-level intelligence because text is a tiny fraction of how humans encode knowledge about the world. A typical four-year-old, he has noted, has absorbed 50 times more data than the world's largest language models not from reading, but from interaction with physical reality. "There is absolutely no way," he said at CES 2025, "that autoregressive LLMs, the type that we know today, will reach human intelligence. It's just not going to happen." Separately, asked about the claim that scaling alone could achieve general intelligence: "Some people claim that if we just keep improving our current technology, we'll achieve human-level intelligence. I've always thought that was nonsense." He has described current LLMs as "an off-ramp, a distraction, a dead end" on the path to systems that genuinely understand the world, and added that "scaling is saturating." His new company, based in Paris, is built around a different paradigm entirely, world models that learn from sensor data and physical interaction rather than text prediction.
Geoffrey Hinton, 2018 Turing Award winner and 2024 Nobel Laureate in Physics for his foundational work on neural networks, left Google in 2023 specifically to speak freely about AI risks. He has acknowledged the hard constraint plainly: scaling requires "huge amounts of computing power and huge amounts of data," and the practical problems of doing so at further scale are real. Hinton's broader concern is not that the transformer will fail quietly but that it will succeed dangerously that systems built in this paradigm will become capable enough to pose alignment risks that the field has not solved. At the AI4 conference in Las Vegas in August 2025, speaking to 5,000 industry leaders, he said: "Once an AI develops its own internal objectives that diverge from ours, even slightly, aligning it becomes extraordinarily difficult." He warned that the companies in the room know about these risks and are not acting on them proportionately. The same month, he urged researchers to instill "nurturing instincts" in AI systems because self-modifying code could evade guardrails in ways no current architecture prevents.
Ilya Sutskever, co-founder of OpenAI and one of the key architects of the GPT lineage, left OpenAI in 2024 over concerns about safety being deprioritized in favor of commercialization. In a rare public interview with Dwarkesh Patel in November 2025, he named the paradigm shift directly: "The age of scaling roughly 2020 to 2025 is ending." Pre-training data is finite. The returns from adding more compute to existing approaches are diminishing. "Is the belief that if you just 100x the scale, everything would be transformed? I don't think that's true. So it's back to the age of research again, just with big computers." He described today's models as "deeply jagged" capable of solving competition mathematics problems while failing at simple generalizations that any child handles without effort. His new company, Safe Superintelligence Inc., valued at $32 billion with no product and approximately 20 employees, is explicitly not trying to build a better version of the current transformer paradigm. It is trying to find a different underlying principle for learning.
Andrej Karpathy, co-founder of OpenAI and former head of AI at Tesla, framed the epistemological problem most precisely in his 2025 year-in-review. Current LLMs, he wrote, are not "evolving animals" on a developmental arc. They are "summoned ghosts" entities shaped entirely by text prediction, reward signals, and human preference ratings, whose intelligence has a completely different structure than biological cognition. "Everything about the LLM stack is different: neural architecture, training data, training algorithms, and especially optimization pressure so it should be no surprise that we are getting very different entities in the intelligence space." The result is jagged capability: a system that is simultaneously a genius polymath and cognitively confused, that aces Olympiad-level mathematics and fails at spatial reasoning a child handles without effort. He describes LLMs as having reconstructed the cortex pattern recognition and language generalization while missing the hippocampus, the amygdala, the cerebellum. The parts of the brain that encode memory, instinct, and motor skill. "We've probably recreated cortical tissue," he wrote, "but we're still missing the rest of the brain."
What makes this pattern notable is who is saying it. These are not outside critics. These are the people who built the transformer, who built GPT, who built the deep learning foundations that made the current wave possible. Their collective position by early 2026 is not that the technology is worthless. It is that the architecture that produced GPT-4, Gemini, and Claude is incomplete as a path to the capabilities being promised and that the institutions continuing to invest billions in scaling it already know this.
The Stanford mathematical proof published in 2024 formalizes what these researchers were describing intuitively. Transformer self-attention hits a hard computational ceiling for complex compositional tasks. Circuit complexity theory shows that transformers are upper bounded by a class of problems (uniform TC⁰) that excludes general Boolean formula evaluation and certain arithmetic tasks, for any realistic parameter budget. The model can learn to approximate these tasks through brute-force memorization of patterns. It cannot solve them through the underlying logical operations the way a human mind or a formal reasoning system does.
That ceiling is not a product roadmap item. It is a mathematical constraint.
The Architecture Shift Nobody Is Talking About Loudly Enough
While the failures above were accumulating in production, something quietly significant was happening in research.
For years, the dominant strategy for building more capable AI was simple: train bigger models on more data with more compute. Scaling laws said performance would follow. And it did. GPT-3 to GPT-4 was four orders of magnitude of compute and a massive leap in capability.
Then the returns started shrinking.
While GPT-4 demonstrated a significant leap in benchmark scores compared to its predecessors, subsequent models have shown only incremental gains, approaching but barely exceeding human-level performance on narrow evaluations. The easy gains from simply scaling model size have been largely exhausted.
Epoch AI estimates the effective stock of quality, repetition-adjusted human-generated public text for AI training at around 300 trillion tokens. If trends continue, language models will fully utilize this stock between 2026 and 2032, or even earlier if intensely overtrained. This is the data wall. The web corpus that fed the current generation of foundation models is running out. Without new fuel, model performance on messy real-world use cases flatlines and risks model collapse, where models remix their own past outputs in a degrading recursive loop.
These are not incremental improvements. They are signals that the architecture built on pure scaling is at a structural limit, and that the next phase will be decided by ideas, not gigawatts.
Nvidia Is Not Just Selling Chips. It Is Financing the Machine.
The story told about Nvidia is one of hardware dominance. The real story is more interesting, and more circular.
In the year ChatGPT first debuted, Nvidia made 16 investments in other companies. In 2024 that number rose to 41. By the end of 2025, it had already made 51 more, not counting a commitment to OpenAI reported at $100 billion. The pattern is consistent: Nvidia invests in a startup, the startup uses the money to buy Nvidia GPUs, and Nvidia books the sale.
The most documented example is CoreWeave. Nvidia invested early, CoreWeave bought Nvidia GPUs at scale, CoreWeave rented those GPUs to OpenAI and Microsoft, then Nvidia committed to purchasing up to $6.3 billion in unsold CoreWeave cloud capacity by 2032 as a financial backstop. When CoreWeave went public in March 2025, Nvidia owned about 7% of it. After a further $2 billion investment in January 2026, that stake climbed to approximately 13%. OpenAI's total commitment to CoreWeave stands at $22.4 billion across three tranches. Microsoft is CoreWeave's largest customer, accounting for 67% of its revenue.
Jay Goldberg, analyst at Seaport Global Securities, describes these deals as a company asking their parents to co-sign their mortgage. Nvidia's equity backing does not just provide capital. It provides credibility that enables CoreWeave to access debt financing at significantly lower interest rates than it could command alone. The chips serve as collateral. Private credit firms Blackstone, PIMCO, Carlyle, BlackRock, are active lenders in what has become a $10 billion-plus debt market for "neoclouds" backed by GPU assets.
The loop is complete and self-reinforcing. Nvidia sells chips. The buyers finance the purchase using the chips as collateral. Nvidia invests in the buyers, raising their creditworthiness. The buyers use the investment to buy more chips. Nvidia buys back computing time from the buyers. The whole structure is capitalized against the continued appreciation of GPU assets. Those assets appreciate as long as demand continues to grow. Demand continues to grow as long as the narrative holds.
Nvidia's portfolio extends well beyond CoreWeave and OpenAI. It holds stakes in Arm Holdings, Applied Digital, Nebius Group, Recursion Pharmaceuticals, Lambda Labs, Scale AI, Intel ($5 billion), UK cloud provider Nscale, and dozens of AI startups. These are not passive financial positions. They are strategic anchors. A company backed by Nvidia gets preferential GPU allocation during supply constraints, when every competitor is on a waiting list.
Jensen Huang calls himself "the chief revenue destroyer" at GTC 2025, deliberately obsoleting each generation before the next ships. Blackwell in 2025. Vera Rubin in Q3 2026. Rubin Ultra in H2 2027. A customer who spent $3 million on a rack this year is on a treadmill. Stop running and you fall behind any lab that did not stop.
The deepest moat is not the hardware. It is CUDA. Nearly 20 years of the CUDA software ecosystem, 4 million developers, 3,000 optimized applications, deep integration into every major AI framework at the level of individual kernel operations. Even as Google, Amazon, Microsoft, and Meta build custom silicon for their own internal inference workloads, they keep buying Nvidia for training and for third-party enterprise customers where they cannot control the software stack. Jim Keller architect behind AMD Zen and Apple's A-series said it plainly: CUDA is not a moat. It is a swamp. Switching costs are so high that even technically superior alternatives cannot displace it without rebuilding the entire ecosystem.
The Memory Oligopoly Nobody Describes as One
Behind Nvidia sits a three-company structure that controls the hardware AI runs on.
SK Hynix, Samsung, and Micron are the only manufacturers in the world capable of producing High Bandwidth Memory at scale. HBM is the specialized stacked memory that sits directly adjacent to the GPU die, feeding it data fast enough to prevent idle cycles. Without sufficient HBM, the most advanced AI processor runs at a fraction of rated performance. This is the memory wall.
SK Hynix dominates with 62% market share. Nvidia accounts for approximately 90% of their HBM supply. SK Hynix confirmed in October 2025 that its HBM, DRAM, and NAND capacity is "essentially sold out" for all of 2026. It has signed supply agreements with OpenAI for the Stargate supercomputer project, which alone is expected to more than double the industry's total HBM requirements.
Micron, the only US-listed pure-play memory company, is posting numbers that have not been seen in the memory industry's history. Fiscal Q1 2026: $13.64 billion in revenue, a 57% year-over-year increase, with gross margins above 50% double what they were in fiscal 2024. For Q2 2026 it guided to $18.7 billion revenue with an approximately 68% gross margin. The entire 2026 HBM supply is already sold out under fixed-price contracts.
For context: the memory industry has historically been one of the most brutally competitive commodity businesses in semiconductors. Margins in the 20–30% range were considered strong. 68% gross margin is not a memory number. It is a pharmaceutical patent number. It reflects a supply-constrained oligopoly extracting rent from a captive market.
HBM production consumes approximately three times the wafer capacity of standard DRAM per gigabyte. Every HBM die fabricated is three ordinary DRAM dies that do not get made. The reallocation is simultaneously starving consumer and enterprise memory markets. Samsung raised prices for 32GB DDR5 modules 60% in a single month in late 2025. Contract DDR5 pricing rose more than 100% over the year. Smartphone production forecasts were revised downward seven consecutive times between October and December 2025, ending at a 7% annual decline. Memory prices increased 246% in 2025 alone.
Samsung's memory division posted 250% year-over-year profit growth in Q4 2025. SK Hynix's operating profit margin in the same period was nearly 56%.
This three-company oligopoly is maximizing profit during a period of unprecedented and inelastic demand, and it is being actively protected.
China has two serious memory manufacturers. YMTC produces NAND and was, until recently, genuinely competitive climbing from less than 1% global market share in 2020 to over 5% by 2023 through genuine technical innovation: 300+ layer Xtacking stacks, first adoption of hybrid bonding in NAND. Then the US put YMTC on the entity list. Tool vendors withdrew. Equipment maintenance stopped. New fab construction halted. By Q2 2025, YMTC's market share had fallen below 5% and was still declining.
CXMT, China's leading DRAM manufacturer, is preparing a $4.2 billion IPO in the first half of 2026. It has demonstrated DDR5-8000 capable chips. It is two to three generations behind on HBM, but in commodity DRAM it is closing the gap faster than the incumbents would prefer.
In March 2026, the Federal Acquisition Regulatory Council issued proposed rulemaking that would ban the use of chips from CXMT, YMTC, and SMIC in government products, with potential extension to commercial supply chains. The stated justification is national security. The functional effect is to wall off the Chinese memory industry from the global supply chain at the precise moment when an alternative supplier could compete on commodity products and relieve the supply pressure driving consumer price increases worldwide.
Micron is the only US-listed pure-play memory manufacturer. It is also building fabs in Idaho and New York under the CHIPS and Science Act with substantial government subsidy. The US government has a financial interest in Micron's profitability that is structurally distinct from any national security calculation. Whether the CXMT rulemaking is primarily security policy or primarily competition policy is a question worth asking out loud.
The Fossil Fuel Policy and What It Is Actually Doing
On January 20, 2025, President Trump declared a national energy emergency and signed executive orders to unleash oil and gas development on federal lands. On April 8, 2025, he signed an executive order to reinvigorate the coal industry, directing agencies to identify coal-powered infrastructure suitable for AI data centers, reclassifying coal as a critical mineral, and making $200 billion in low-cost DOE financing available for coal energy investments.
On July 23, 2025, a further executive order accelerated federal permitting specifically for AI data centers requiring over 100 MW of new load. Approved power sources: natural gas turbines, coal power equipment, nuclear power equipment, geothermal. Solar and wind projects face additional approval burdens. Interior Department staff require the Secretary's personal sign-off on renewable energy siting on federal land.
Simultaneously, the Interior Secretary issued Order No. 3437, identifying solar and wind as "expensive, unreliable, foreign-controlled intermittent energy sources" and directing staff to eliminate any guidance designed to facilitate their development on public lands.
The administration claims 74 coal plants saved from regulatory closure. DOE identified four federal sites for AI data center development: Idaho National Laboratory, Oak Ridge Reservation, Paducah Gaseous Diffusion Plant, and Savannah River Site.
The technical argument for fossil fuels in AI infrastructure has a real kernel. Data centers need dispatchable baseload power that matches their consumption profiles around the clock. Renewable energy without massive grid-scale storage does create reliability challenges at the power densities AI infrastructure demands. This is a genuine engineering constraint.
The policy response, however, goes substantially beyond solving that engineering problem.
Reclassifying coal as a critical mineral. Making $200 billion in federal credit available for coal infrastructure. Requiring personal Secretary approval for renewable energy permitting on 530 million acres of federal land while fast-tracking fossil fuel projects. These are not responses to a grid reliability calculation. They are industrial policy decisions that redirect federal financing and land access toward fossil fuel industries under cover of AI urgency.
In March 2026, the White House announced the Ratepayer Protection Pledge, bringing Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI together to commit to build, bring, or buy new generation resources for their data centers. The companies signing a pledge not to harm electricity ratepayers are the same companies planning 150 to 200 gigawatts of data center capacity through 2030, capacity that will require energy generation roughly equivalent to adding a second France to the US grid.
Goldman Sachs found AI contributed "basically zero" to US GDP growth in 2025. 80% of firms using AI report no measurable productivity or employment gains. The infrastructure being accelerated on federal lands carries no such accountability condition. There is no mechanism that ties federal land access, federal permitting priority, or federal financing to demonstrated economic output.
The communities absorbing the environmental costs are not the communities holding the equity. The households seeing electricity rate increases from data center demand are not the households holding Nvidia stock. The aquifers being drawn down for server cooling do not appear in any balance sheet.
The energy policy, the semiconductor policy, the investment structure, and the narrative are all oriented in the same direction. They are oriented toward continuation of the current regime regardless of whether the current regime is producing the returns it advertises.
That is not an engineering outcome. It is a political economy outcome.
90 Decibels, Permanently
In Northern Virginia's Loudoun County, a resident who moved there for the forested areas and farmland now cannot open his windows on cool evenings. The noise from the gas turbines powering a nearby data center registers off the charts on decibel reader apps, sometimes reaching 90 decibels near houses. A neighbor put mattresses against the windows to block it.
This is not a localized complaint. It is the pattern.
Data centers generate significant noise pollution primarily from diesel generators and HVAC systems, with internal noise levels reaching up to 96 decibels, well above the 85 decibel threshold considered harmful to hearing over time. This is persistent, continuous, and it affects data center workers, nearby residents, and local wildlife.
Communities from rural Pennsylvania to suburban Arizona to historically Black neighborhoods in Memphis are pushing back against projects that bring few long-term jobs but outsized impacts: spiking electricity demands, strained water supplies, noise pollution, and the imposition of high-voltage transmission lines through residential and agricultural areas.
An analysis of communities within one mile of EPA-regulated data centers found they are disproportionately communities of color relative to the national median, and they face particulate matter, nitrogen dioxide, and diesel particulate matter levels above the national median. About 4 million people live within one mile of these 550 tracked data centers. The universe of untracked and unpermitted data centers is larger.
Black Americans already suffer disproportionately from air pollution and other environmental injustices. Low-income Black Americans have the highest mortality rate due to fine particulate matter air pollution of any demographic group in the country. Data centers with diesel backup generators emit 200 to 600 times more nitrogen oxides than a natural gas plant generating the same amount of energy. The IEEE estimated that data centers caused about $6 billion in public health damages from air pollution alone in 2023.
A 2025 model published in Eco-Environment & Health estimates that US data centers could contribute to nearly 1,300 deaths annually by 2030, resulting in a public health burden of more than $20 billion per year.
The NAACP held a two-day summit in December 2025 under the theme "Stop Dirty Data," with leaders calling out what they described as a familiar playbook: heavy industrial facilities sited near communities of color, with nondisclosure agreements preventing residents from knowing what is being emitted, what water is being consumed, or where the wastewater goes.
"There are still some people who are not connecting their ChatGPT requests or the AI mode that shows up with Google to actually being an infrastructure that lives in a community," said the NAACP's director of environmental and climate justice.
The Trump administration's rollback of Project 2025's assertion that race should not be considered in environmental justice review compounds the exposure. There is no federal framework left to challenge siting decisions that disproportionately burden communities of color.
Training's Next Phase: Synthetic Data and Its Ceiling
Given that the web is running out of quality training data, every major lab has turned to synthetic data: AI generating training examples for future AI.
The appeal is obvious. You don't need the open internet. You don't need to pay content licensing fees. You can generate targeted, domain-specific training examples at near-zero marginal cost. DeepSeek's R1 showed that reinforcement learning alone, training a model by having it play against itself and learn from what works, could produce reasoning abilities comparable to supervised methods. AlphaGo Zero did this for the game of Go. The model that beat the world champion trained on zero human games.
But language is not Go. Synthetic data scales human judgment. It does not replace it. The underlying corpus must remain human to provide context and prevent model drift.
The boundary condition is called model collapse. When models are fine-tuned with data they synthesized themselves, it takes only a few rounds for the model to start producing degraded output. The collapse is irreversible. You can generate a billion synthetic reasoning problems. But if they were all generated by a model that was already slightly wrong, the next model learns the wrongness. And the next. Until the signal becomes noise.
The next wave of AI progress will be decided less by bigger models and more by who controls clean human data. The public web is running dry. Who controls what gets quietly excluded will determine whether AI stays grounded or drifts into self-trained sameness.
This is not a problem that compute can fix. It is a data provenance problem. And the data provenance problem cannot be solved by spending more.
The Accountability Gap
Here is the ground-level reality of AI software in 2026, stated as flatly as the data allows.
Multiple independent research studies confirm accelerating code quality degradation, exponential security vulnerability growth, and unsustainable maintenance burdens. The data comes from GitClear's analysis of 153 million lines of code, Carnegie Mellon's study of 800+ GitHub repositories, Apiiro's Fortune 50 security research, and Stack Overflow's 2025 Developer Survey. The data tells a consistent story.
Stack Overflow's survey found 84% of developers now use or plan to use AI tools, up from 76% the year before. But trust in AI accuracy has declined. 45% of developers cite that AI solutions are "almost, but not quite, right" as their primary frustration. These near-misses lead to subtle bugs that are time-intensive to detect and resolve. Over one-third of Stack Overflow visits now stem from issues related to AI-generated code.
Goldman Sachs found no meaningful relationship between AI and productivity at the economy-wide level. But they found a median productivity gain of around 30% for two specific, localized use cases. In targeted functions, the technology is already delivering on its transformative promises.
That gap, no macro effect but real micro gains in narrow domains, is the honest picture of where AI is in early 2026. Not a revolution. Not a failure. A technology that works well in some specific contexts and is being deployed everywhere regardless.
Far from isolated edge cases, failures surface in nearly every major sector. The wider and faster AI scales, the higher the stakes of every miscalculation baked into its code or data.
The companies winning are not the ones spending the most. They are the ones being precise about where the 30% gain applies, not treating it as a universal law.
The Map Versus the Territory
The map says: AI is a transformative technology delivering productivity gains across the economy, powered by continuous scaling improvements, with hardware capacity growing to match every demand.
The territory says: most deployments fail or are abandoned. The ones that reach production often contain critical vulnerabilities. The scaling strategy is hitting data and architecture limits simultaneously. The hardware supply chain is sold out through 2026 and causing component shortages across consumer electronics. The grid serving these data centers is already in reliability deficit in multiple major regions, and that deficit widens through 2029. The people who built the transformer the actual inventors, the researchers who gave the industry this decade of capability gains have left the labs or stated publicly that the current architecture is insufficient for what is being promised. The economic gains are real but narrow and highly localized. The energy infrastructure being built to support current-generation models will outlast the models themselves.
None of this means the technology is not real. The capability gains from reinforcement learning on verifiable rewards are genuine. The narrow productivity improvements Goldman found are genuine. The models performing at gold-medal level on mathematics competition problems are genuine.
What is not genuine is the version of the story where all of this is already working at scale, already contributing to broad economic growth, already ready for autonomous deployment without human oversight, already worth the grid strain and rate increases and aquifer depletion and community health burden being imposed now, before the productivity case has been demonstrated at the scale that would justify them.
The infrastructure is being built as if the map is correct. The map is being maintained by the people most invested in not updating it. The territory is where the rest of us live.
The real question is not whether you believe in the technology. It is whether you can identify who bears the cost when the productivity proof doesn't arrive, and whether anyone with the power to ask has actually asked them.