AI Is Making These Things More Expensive
From chips to nuclear plants, the world's resources are being reallocated around AI. This isn't a prediction. It's happening now.
The World No Longer Revolves Around Humans
The world no longer revolves around humans.
Or more precisely: capital no longer makes money by serving humans.
Capitalists used to ask: How can I get more people to buy my products? How can I hire more people to work?
Now they ask: How can I make AI run faster? How can I feed AI more data? How can I keep AI working 24/7?
Humans serve capital. Capital serves AI.
Therefore, humans serve AI.
This is the economic reality of 2026.
Below is a list of things whose prices are rising because of AI.
AI Needs a Dedicated Computer
I recently reformatted all my idle Windows machines to Linux. My AI needs them.
Many experts recommend Mac Mini. It's cost-effective.
Clawdbot just went viral. It connects AI to social apps, lets you assign tasks, and keeps AI on standby 24/7 to manage calendars, write code, reply to emails, and monitor inboxes.
AI needs a machine that never shuts down.
This AI tool is directly driving Mac Mini sales.
Then there's the "Ralph Wiggum" technique: make AI loop through tasks until conditions are met.
Software tools that support continuous AI operation are proliferating.
HBM Is the Real Chip Bottleneck
GPU price hikes are old news. The NVIDIA RTX 5090 launched at $1,999 but trades at $4,499.
But GPUs are just the tip of the iceberg.
HBM (High Bandwidth Memory) is the real bottleneck.
HBM is the "blood vessel" of AI chips, responsible for high-speed data transmission. Without HBM, even the most powerful GPU can't run.
HBM production requires TSV stacking and advanced packaging. Yields are low. Capacity is limited.
Only three companies can make it: SK Hynix (60% market share), Samsung, and Micron.
SK Hynix has announced: HBM capacity for 2024, 2025, and 2026 is sold out.
Order now, and you won't receive delivery until 2027 at the earliest.
Your RAM Is AI's Leftovers
The HBM shortage has triggered a chain reaction.
Manufacturers are shifting capacity to HBM. Consumer DRAM has become "leftovers."
DDR5 price changes:
- September 2025: 16GB modules at $7-8
- November 2025: $13
- December 2025: $27
Nearly 300% increase in three months.
Samsung raised DDR5 contract prices by over 100% between November and December 2025.
Analysts say: don't expect prices to drop before 2027. Peak may arrive mid-2026.
The memory in your PC and the HBM in AI data centers come from the same wafer.
Manufacturers chose the more profitable one.
ABF Film: The Hidden Monopoly
ABF (Ajinomoto Build-up Film) is an insulation material used in chip packaging. CPUs, GPUs, and AI accelerators all need it.
95% of the global ABF market is monopolized by one company: Ajinomoto Fine-Techno.
Yes, the same Ajinomoto that makes MSG.
AI chips require more ABF layers, finer line widths, and stricter warpage control.
Each AI accelerator uses several times more ABF layers than traditional SoCs.
But ABF capacity expands slowly. The manufacturing process is complex and yield-sensitive.
Result: CoWoS advanced packaging lead times now exceed 12 months.
ABF capacity will remain tight through 2026.
NVIDIA Has Locked Up Laser Chips
800G optical modules need EML (Electroabsorption Modulated Laser) chips for data transmission.
This is a critical component for high-speed interconnects within AI data centers.
NVIDIA has locked up most of the capacity from Lumentum, the main EML supplier.
Lead times extend beyond 2027.
In Q3 2025, EML laser chip supply gaps reached 25-30%.
Google, Meta, and Amazon AWS are scrambling for alternatives like CW lasers. But they can't fully replace EML in the short term.
Not just chips are in shortage. The "cables" between chips are too.
Glass Is the New Bottleneck
Speaking of cables: fiber optics are also in shortage.
A major fiber optic manufacturer has announced: inventory is sold out through 2026.
Ribbon fiber lead times exceed 60 weeks. Loose-tube fiber delivery is scheduled for Q3 2026.
Why?
AI data centers require 16 to 36 times more fiber than traditional CPU data centers.
Glass, yes, the raw material for fiber optics, has become a new bottleneck.
Rare Earths: China Controls the Choke Point
Neodymium magnets are used in data center cooling systems, power supplies, storage systems, and even GPUs and TPUs.
The supply gap for neodymium and praseodymium is expected to expand to 22% of supply by 2030.
The global rare earth supply chain is highly concentrated in China. Mining, processing, magnet production, almost entirely in China.
In 2025, China imposed export controls on rare earths. The semiconductor and AI industries were directly affected.
The US is scrambling to build a non-China rare earth supply chain.
But this will take years.
3M Is Discontinuing Coolant
AI servers generate too much heat. Traditional air conditioning isn't enough.
Servers need to be immersed in liquid for cooling. This is called immersion cooling.
3M's Fluorinert is the leading dielectric coolant.
But Fluorinert is a PFAS ("forever chemical"). Environmental regulations are tightening.
3M announced: Fluorinert production ends in late 2025.
Order deadline was March 31, 2025. Miss it and it's gone.
Fluorinert prices now exceed $200/liter.
Alternatives are emerging. InnoChill sells for $15-25/liter, over 80% cheaper.
But ecosystem transitions take time.
The immersion cooling market was $5.4 billion in 2024. It's projected to reach $66 billion by 2034.
AI's heat is spawning a new multi-billion dollar industry.
The Power War
Global data center electricity consumption in 2024: 415 TWh, 1.5% of global electricity consumption.
Expected to double by 2030.
US data center electricity consumption in 2024: 183 TWh, over 4% of national consumption. Expected to grow 133% to 426 TWh by 2030.
A single ChatGPT query consumes roughly 5-10 times more electricity than a regular Google search.
Virginia data centers currently consume 25% of the state's electricity. By 2030: 46%.
So tech giants are restarting nuclear plants:
- Microsoft signed a 20-year power purchase agreement to restart Three Mile Island
- Google became the first to sign with a small modular reactor
- Amazon is investing in small nuclear reactors
AI has revived nuclear power.
Transformers: 3-6 Year Wait
How does electricity reach data centers? Transformers.
Transformer lead times have extended to 3-6 years.
Before 2019, delivery took only a few months.
The problems: material shortages, outdated manufacturing processes, frequent extreme weather.
North America has over 80,000 different transformer configurations. Standardization is impossible.
US transformer supply gap expected to reach 30% in 2026.
Even with money to build a data center, you might not get enough transformers.
Copper: A Decade-Long Gap
Transformers need copper. Cables need copper. Cooling systems need copper.
A large data center can consume over 2,000 tons of copper.
In Q1 2025, global copper demand grew 3.3% YoY.
Over the next decade, data centers are expected to consume an average of 400,000 tons of copper annually, peaking at 572,000 tons in 2028.
S&P Global forecasts: copper demand will grow from 28 million tons in 2025 to 42 million tons by 2040.
Without significant supply expansion, there could be a 10 million ton shortage by 2040.
The copper market is expected to be short 124,000 tons in 2025 and 150,000 tons in 2026.
AI industry growth is expected to increase global copper demand by over 15% in 2025.
Land Prices Double
Industrial land in Loudoun County, Northern Virginia: over $4 million per acre in 2025.
Price increases:
- Loudoun County: 45%
- Prince William County: 38%
- Some areas near Leesburg: doubled
Land scarcity in data center corridors is pushing developers toward multi-story buildings and secondary counties.
Texas is expected to become the largest US data center market within three years.
Central Texas data center construction grew 4x from 2023 to 2024.
The "Stargate Project," announced in January 2025, plans to build up to 20 AI data centers across the US. Texas is the core hub.
AI isn't just competing for chips. It's competing for land.
Natural Gas Demand Surges
AI data centers could increase US natural gas demand by 200-300 million cubic feet per day between 2025 and 2027.
That's 3% of total US consumption.
By 2030, AI-related natural gas demand could reach 500-800 million cubic feet per day.
Why? The grid can't keep up. Data centers need their own power plants. Natural gas plants are cleaner and more stable than diesel generators.
Chevron and GE Vernova are building dedicated natural gas plants for data centers.
AI is driving traditional energy demand.
The Complete Price Chain
Let's connect the dots:
AI training → needs GPUs → GPUs need HBM → HBM crowds out DRAM capacity → memory prices rise
GPUs need packaging → packaging needs ABF film → ABF supply tightens → chip delivery delays
GPU interconnects need optical modules → optical modules need EML lasers → NVIDIA locks up capacity → 25-30% supply gap
Optical modules need fiber → fiber needs glass → 60+ week lead times
Servers need cooling → cooling needs dielectric fluid → 3M discontinues production → prices surge
Data centers need power → power needs the grid → the grid needs transformers → 3-6 year lead times
Power transmission needs copper → copper supply gap widens → potential 10 million ton shortage by 2040
Data centers need land → Northern Virginia land prices double
Grid can't keep up → need private power plants → natural gas demand increases 3%
This is the new economic logic of 2026.
Why This Matters to You
You might think this doesn't affect you.
Wrong.
- Electricity rates are rising. Some Virginia areas saw 24% increases last year.
- Memory is rising. Want to upgrade your PC? Prices have multiplied.
- Cloud services are rising. Enterprise AI spending went from $63,000/month in 2024 to $85,000/month in 2025.
Resource allocation priorities are being reshuffled:
- Power: Electricity that should go to residential use increasingly flows to data centers
- Memory: DRAM that should go to consumer electronics is being crowded out by HBM
- Land: Plots that should be for housing and commercial use become data centers
- Copper: Copper that should go to infrastructure flows to tech companies
AI's priority is surpassing many traditional needs.
What's Permanently Scarce vs Temporarily Scarce
Not all shortages are created equal.
Some are structural, meaning supply-demand dynamics have been permanently altered. Others are temporary capacity bottlenecks that will ease in a few years.
Structural Scarcity (Long-term):
Copper. New mine development takes 15+ years. Ore grades keep declining. Extraction costs keep rising. Potential 10 million ton gap by 2040. J.P. Morgan forecasts 2026 will have the largest supply gap in 22 years.
Rare Earths. 90% of mining and processing is concentrated in China. Western supply chain rebuilding will take 5-10 years minimum. China began export controls in 2025.
Power Infrastructure. The 3-6 year transformer lead time isn't due to demand surge. It's because industry capacity was already insufficient and standardization is low. This is structural.
HBM Capacity Allocation. IDC analysts say this isn't a temporary shortage. It's a "permanent strategic reallocation of global silicon wafer capacity." By 2026, data centers will consume 70% of global memory capacity.
Temporary Scarcity (Capacity Catching Up):
EML Laser Chips. Q3 2025 gap of 25-30%, but capacity is expanding. Expected to ease in H2 2026.
ABF Film. Ajinomoto and others are expanding. Some new factories expected at full capacity by 2026.
HBM Pricing. Though capacity is permanently allocated to AI, SK Hynix, Samsung, and Micron are all expanding. H2 2026 may shift from seller's to buyer's market.
DDR5 Pricing. Expected to peak mid-2026, then decline starting 2027.
Fiber. Manufacturers are expanding capacity. The gap is mainly due to demand growing too fast, not production constraints.
Alternative Technologies Are Coming
Supply chain pressure is spawning alternative technologies.
Glass Substrates Replacing ABF Film
Intel and AMD are both advancing glass substrate technology. In January 2026, Intel announced its first commercial product using a glass core: Xeon 6+ "Clearwater Forest" entering mass production.
Glass substrate advantages: better flatness, less warpage, 10x interconnect density improvement, 60% lower dielectric loss.
More importantly, glass is cheaper than silicon interposers and can be made in larger panels, reducing packaging costs.
Silicon Photonics and Co-Packaged Optics
2026 is the commercialization year for silicon photonics.
NVIDIA's Spectrum X Photonics and Quantum X Photonics switches already use co-packaged optics (CPO). Broadcom's Tomahawk 6 switch ships in 2026.
CPO can reduce data transmission energy consumption by over 70%. Silicon photonics penetration in the high-end 1.6T market is expected to reach 50-70%.
This will ease EML laser chip shortages. Google, Meta, and AWS are shifting to continuous wave (CW) lasers as alternatives.
Custom ASICs Replacing General-Purpose GPUs
NVIDIA holds 85% market share, but the landscape is changing.
In 2026, custom ASIC shipments from cloud providers are expected to grow 44.6%, far exceeding GPU shipment growth of 16.1%.
Google has TPUs. Anthropic uses over a million of them. Amazon has Trainium and Inferentia. Microsoft has Maia 200. Broadcom customizes XPUs for Meta, Google, and ByteDance.
In January 2026, NVIDIA acquired Groq for $20 billion, integrating its low-latency LPU technology into the next-generation Vera Rubin platform.
This means: NVIDIA itself is playing defense, because specialized chips are more power-efficient and cheaper for inference workloads.
Software Optimization Reducing Inference Costs
Over the past three years, inference cost per query has dropped 1,000x.
Key techniques include: model quantization, knowledge distillation, semantic caching, automatic routing.
Semantic caching can save 30-70% on repetitive queries. Smart routing can direct load to the smallest viable model, saving 40-60%. LLM cascade routing can save up to 98%.
NVIDIA's Rubin platform promises 10x reduction in inference token cost and 4x reduction in GPUs needed for MoE models.
Where Are the Opportunities?
Having covered scarcity, let's discuss opportunities.
Copper Mining Stocks
Copper prices rose 41-44% in 2025. Q2 2026 forecast: $12,500/ton. J.P. Morgan's full-year average forecast: $12,075/ton.
Low-cost, expandable, financially healthy copper miners are direct beneficiaries. Junior copper miners outperformed in 2025.
Rare Earths
The MVIS Global Rare Earth/Strategic Metals Index rose over 60% in the past year.
Since January 2025, the US, Canada, Australia, Japan, and Europe have invested over $3 billion in rare earth supply chains. The US alone: $1.4 billion.
Opportunities span mining, processing, and magnet manufacturing.
Nuclear Power
Meta signed a 20-year, 2,600 MW power purchase agreement with Vistra. Signed with Oklo to prepay for a 1.2 GW Aurora powerhouse.
Amazon invested $500 million in X-Energy, planning over 5 GW of nuclear capacity. Google signed with Kairos Power for 500 MW.
Constellation Energy (the largest US nuclear operator) has signed with Meta and Microsoft.
But be careful: SMR startups like Oklo and NuScale have soaring stock prices but nearly zero revenue and high cash burn.
Utilities
Data centers are driving sustained electricity demand growth. American Electric Power plans to invest $72 billion from 2026-2030. Data centers account for 80% of incremental load.
NextEra Energy, Duke Energy, Xcel Energy, and Dominion Energy are all increasing capital expenditure.
Utilities are the "picks and shovels" of the AI boom.
Data Center REITs
Equinix and Digital Realty benefit from AI demand, but valuations are no longer cheap.
Applied Digital (transitioned from crypto mining) has higher growth potential.
Cooling Equipment
Vertiv and Eaton provide data center cooling and power equipment. The immersion cooling market: $5.4 billion in 2024, projected $66 billion by 2034.
Optical Communications
Micron (HBM), Credo Technology (high-speed connectivity), Ciena (optical networking), and Arista Networks (network switches) are all direct beneficiaries of AI infrastructure.
Semiconductor Equipment and Materials
Glass substrates, silicon photonics, advanced packaging.
Equipment and materials suppliers in these areas will benefit from technological transitions.
Risk Warnings
Finally, some risks.
AI Bubble Risk. If AI commercialization falls short of expectations, demand across the entire chain will contract.
Geopolitical Risk. Taiwan (TSMC) and China (rare earths) are single points of failure. Any conflict would rupture supply chains.
Technology Substitution Risk. If an alternative technology matures faster than expected (optical computing, neuromorphic chips), current bottlenecks could suddenly disappear.
Overinvestment Risk. When everyone is expanding capacity, oversupply could arrive faster than expected. HBM is an example: H2 2026 may shift from shortage to oversupply.
Regulatory Risk. Environmental regulations (PFAS coolants), antitrust (NVIDIA), and export controls (rare earths, chips) could all change the landscape.
Closing Thoughts
I'm not saying this is good or bad.
I'm simply describing what is happening.
Capital is profit-seeking. Where returns are highest, capital flows.
Right now, the highest returns are in AI infrastructure.
And so, the world's resource allocation is being rearranged around AI.
Author
Categories
More Posts
Google Antigravity: The Most Cost-Effective Vibe Coding Tool in 2025
An in-depth hands-on review of Google Antigravity's Pro Plan. Discover why it's a great value AI coding tool, featuring Claude Opus 4.5, deep web search, Nano Banana image generation, and multi-agent workflows.
苏江：Gemini是当今最好的决策辅助工具
苏江：Gemini是当今最好的决策辅助工具
共识的尽头
当AI学会欺骗、算法可以证明信任、智能合约自动执行契约——人类文明建立在真实、语言与共识三个不稳定基础上的信任体系，正在被重新定义。AI时代认知三部曲之三。
Need a Custom Solution?
Still stuck or want someone to handle the heavy lifting? Send me a quick message. I reply to every inquiry within 24 hours—and yes, simple advice is always free.
Newsletter
Join the community
Subscribe to our newsletter for the latest news and updates