2025 in Review

The year that AI's constraints evolved faster than anyone expected

Author

Justin Wright
January 05, 2026 • Est. Reading Time: 13 minutes

Happy Monday!

Looking back at 2025, one pattern emerges above all others: we kept solving bottlenecks, only to discover bigger ones we never saw coming.

We started the year convinced that scale and compute were everything. We're ending it knowing that efficiency, power access, and infrastructure matter more. We thought GPUs were the constraint. Then it was electricity. Then memory chips. Then geography itself.

2025 taught the AI industry a hard lesson: the bottleneck always moves faster than you think.

Here are the five moments that defined AI's year and reshaped what "competitive advantage" actually means.

📊 THE YEAR IN NUMBERS:

  • DeepSeek R1: $6M to train vs. OpenAI's $100M

  • Nasdaq lost $1T in one day (January 27)

  • Lumentum stock: +372% (Nvidia: +97%)

  • Grid connection queues: 4-7 years in Virginia

  • Data center deals: Record $61B

  • Micron shut down consumer memory business

  • Enterprise AI spend: $11.5B → $37B (3.2x)

  • Build-to-buy flip: 47% built internally (2024) → 24% (2025)

2025 was the year AI's constraints kept moving. DeepSeek proved efficiency beats brute force in January, rattling markets. Power became the real bottleneck as grid queues hit 4-7 years. Memory chips got cannibalized by AI demand, forcing Micron to exit consumer markets. Infrastructure stocks crushed model companies. By year-end, 76% of companies were buying AI instead of building it.

TL;DR

January's $1 Trillion Wake-Up Call

On January 27, 2025, a Chinese startup most people had never heard of sent Silicon Valley into panic mode.

DeepSeek released R1, an open-source reasoning model that matched OpenAI's o1 and outperformed it on certain benchmarks. The kicker? DeepSeek trained it for under $6 million using less powerful H800 GPUs. OpenAI reportedly spent over $100 million on similar capabilities using top-tier chips.

By market open, the Nasdaq had lost $1 trillion in value. Nvidia dropped 17% in a single day. The conventional wisdom, that frontier AI required massive capital and cutting-edge hardware, shattered in real time.

What happened: DeepSeek used Mixture-of-Experts architecture and reinforcement learning without human feedback, drastically reducing training costs. Instead of activating all parameters at once, only relevant "expert" modules woke up for each query. Export controls were supposed to slow China down. They forced Chinese labs to get smarter instead.

Why it mattered: Marc Andreessen called it "AI's Sputnik moment." The assumption that only well-funded Western firms could compete evaporated overnight. The "scale is all you need" dogma died. Efficiency suddenly mattered more than compute budget.

What changed: By February, Alibaba's Qwen and Allen Institute's Tulu both claimed to beat DeepSeek. The efficiency race was on. Companies stopped asking "how big can we scale?" and started asking "how smart can we optimize?" UC Berkeley researchers matched o1 performance for $450 by year-end. After DeepSeek released its approach? That dropped to $50. The message was clear: just throwing money at the problem wasn't a viable strategy anymore.

The Power Wall Nobody Saw Coming

While everyone was still processing DeepSeek, a different constraint was building. Companies had solved the GPU shortage, they just couldn't plug them in.

Microsoft's CEO admitted the company "doesn't have enough electricity to install all the AI GPUs in its inventory." Grid connection queues in Northern Virginia, the world's largest data center market, stretched to 4-7 years. OpenAI had to stagger its GPT-4.5 launch in February because it ran out of power capacity.

What happened: AI data centers require 30-80 kilowatts per rack vs. 8-15 kW for traditional computing. Global data center electricity demand jumped from 415 TWh to a projected 945 TWh by 2030 (nearly 3% of total global consumption). Single training runs could require up to 8 gigawatts by 2030. That's eight nuclear reactors for one model.

The infrastructure didn't exist. Grid upgrades needed $720 billion through 2030. Transmission projects took years to permit and years more to build. The bottleneck wasn't silicon anymore, it was kilowatts.

Why it mattered: Microsoft restarted the Three Mile Island nuclear plant out of desperation. Alphabet paid $4.75 billion for Intersect, an AI data center company, to buy power contracts. Pennsylvania launched a state program guaranteeing electricity before offering tax incentives, explicitly competing with Virginia on energy access. Geographic advantage flipped overnight. Coastal tech hubs with talent meant nothing if you couldn't keep the lights on.

What changed: "Bring Your Own Power" became a real strategy. Companies started deploying on-site generation rather than waiting in grid queues. Energy partnerships became as important as cloud deals. The conversation shifted from "where are the engineers?" to "where can we actually build?"

The Great Memory Cannibalization

By summer, a new shortage emerged. This one hit consumers directly.

In December, Micron Technology, one of the world's top memory chipmakers, shut down its entire consumer memory and SSD business. Not because it was unprofitable, but because AI data centers were willing to pay more.

The company reported being "more than sold out" of memory chips, with the best revenue upside in the history of the U.S. semiconductor industry excluding Nvidia. AI workloads require 2-4x the power of traditional server chips and exponentially more memory. Micron redirected everything to meet that demand.

What happened: AI servers need massive amounts of RAM for training and inference. A single hospital using AI for medical imaging processes 7 billion images. Data centers needed high-bandwidth memory (HBM) that traditional consumer products don't use, but the supply chain overlap meant trade-offs.

Dell's COO admitted in November: "I don't see how this will certainly not make its way into the customer base." Translation: PC and phone prices are going up because AI ate the chip supply.

Why it mattered: This was the first time regular consumers felt a direct cost from the AI boom. Not abstract concerns about jobs or misinformation. Actual price increases on devices they buy. The "who pays for AI?" question moved from theoretical to tangible.

Storage companies like Seagate and Western Digital tripled in value as AI's appetite for hard drives exploded. AI doesn't just need processing power. It needs somewhere to put the data, and that somewhere is cannibalizing consumer supply chains.

What changed: The AI infrastructure boom started showing up in consumer economics. Memory chip shortages. Higher device costs. These weren't externalities anymore, they were showing up in household budgets. The invisible hand of AI's infrastructure demands became visible in your laptop's price tag.

When Infrastructure Crushed AI Darlings

While everyone watched OpenAI's model releases and Nvidia's earnings, a different story was unfolding in the markets.

Lumentum, a company that makes fiber-optic connections for data centers, surged 372% in 2025. Seagate and Western Digital both tripled. Celestica jumped over 300%. Even Micron, despite shutting down consumer business, posted record results.

Nvidia? Up 97%. Still impressive. But the infrastructure plays crushed the chip king.

What happened: Data center deals hit a record $61 billion through November. The largest single transaction: a $40 billion acquisition of Aligned Data Centers by a consortium including BlackRock, Microsoft, and Nvidia. They weren't buying a chip company or an AI lab. They were buying physical capacity with secured power.

Every GPU in a rack needs optical connections to every other GPU. Future systems need rack-to-rack connections, then data-center-to-data-center. Lumentum's CEO said 60% of revenue now comes from AI infrastructure. The company's sales jumped 58% in one quarter.

AI doesn't just need chips. It needs cooling systems, power distribution, storage, networking, and physical buildings that can handle the heat and energy load. The picks-and-shovels play was outperforming the gold rush.

Why it mattered: The market started pricing in a new reality: AI's competitive advantage wasn't about model quality anymore. It was about infrastructure control. Microsoft spent $80 billion on AI data centers in fiscal 2025. Meta, Amazon, and Google collectively poured hundreds of billions into capex as well. But the companies supplying the infrastructure captured even more value.

What changed: The investment thesis shifted. Instead of "who has the best model?" it became "who has guaranteed access to power, memory, and cooling?" The entire supply chain from power generation to fiber optics became strategic. Companies buying data centers for their energy contracts, not their compute, was the signal. Infrastructure wasn't the boring supporting act anymore; it became the main event.

The Build-to-Buy Reversal

By year-end, a quieter but equally significant shift had completed: companies gave up on building AI themselves.

In 2024, the split was almost even: 47% of companies built AI solutions internally, 53% bought them. By 2025, that flipped dramatically: only 24% still built, while 76% bought.

Bloomberg trained BloombergGPT in 2022. Walmart built Wallaby in 2024. Everyone thought custom models were the future. Then everyone realized the moat wasn't the model, it was the application layer and the data.

What happened: Enterprise AI spending jumped from $11.5 billion to $37 billion, a 3.2x increase. But over half went to applications, not infrastructure. Vertical AI like legal, healthcare, and government tripled to $3.5 billion because these were workflows companies couldn't build themselves.

Only 48% of AI initiatives made it from pilot to production, taking an average of 8 months. The 95% of companies getting "zero return" from AI (per MIT's July study) weren't failing because the technology didn't work. They were failing because building and maintaining it was harder than expected.

Why it mattered: The "every company will build proprietary AI" narrative from 2023-2024 died quietly. Companies realized that buying best-in-class applications and focusing on integration delivered better ROI than building from scratch. The competitive advantage wasn't owning the model, it was deploying it effectively.

Product-led growth worked for AI in ways it never did for traditional enterprise software. Developers could try tools immediately, companies could prove value fast, and the switching costs were lower than expected.

What changed: The window for horizontal AI infrastructure plays narrowed significantly. Winners were either vertical specialists (solving specific industry problems) or infrastructure providers (offering guaranteed power, compute, and deployment). The vast middle of "build your own AI" became untenable for most companies.

By December, the market had clarity: build if you're a tech giant with billions to spend, buy if you're everyone else, and focus your resources on the application layer where differentiation actually matters.

Lessons for 2026

The next bottleneck is already here, we just don't see it yet.
We solved GPUs, hit power, solved power (somewhat), hit memory. The pattern is clear: fix one constraint, expose the next. In 2026, watch for network bandwidth, cooling system capacity, or skilled operators becoming the new choke points.

Constraints force better innovation than abundance.
DeepSeek proved it: export controls didn't slow China down, they forced efficiency innovation that ultimately advanced the entire field. The companies that win in 2026 won't be the ones with unlimited budgets. They'll be the ones who optimize around current constraints.

Geography matters more than talent for infrastructure.
Pennsylvania competing on guaranteed power, not tax breaks or talent pools, signals a fundamental shift. In 2026, expect more states and countries to compete on energy access and permitting speed. Where you can build matters more than where people want to live.

Infrastructure captures more value than models.
Lumentum +372%. Nvidia +97%. The lesson is clear: in 2026, pay attention to who's supplying power, memory, cooling, and connections, not just who's releasing the next model. The enabling layer is where the money flows.

Efficiency is the new scale.
The era of "bigger is better" is over. UC Berkeley matching o1 for $50, down from $100 million, shows where things are heading. In 2026, watch for models that do more with less, not just more with more.

In motion,
Justin Wright

If 2025 taught us that every constraint we solve just reveals the next one, and if the pattern keeps accelerating, are we approaching a point where the bottlenecks compound faster than solutions can scale? Or will distributed architectures and efficiency gains finally break the cycle?

Food for Thought

I am excited to officially announce the launch of my podcast Mostly Humans: An AI and business podcast for everyone!

Episodes can be found below - please like, subscribe, and comment!