Author's Note: This piece has been sitting in my drafts for weeks. I started it, got halfway through, then Mind's Eye broke, then Ubisoft happened, and this just... collected dust. It's functional now. The data's current. But I've been staring at it long enough that I'd rather publish it imperfect than revise it into a sixth draft nobody asked for.

When I started this investigation, I expected conspiracy theories. What I found was something more mundane and more alarming: documented market concentration so extreme it makes Standard Oil look competitive.

The tech industry isn't secretly manipulated. It's structured as a series of oligopolies, and AI demand just stress-tested every one of them to failure.

The Numbers

Let's start with what's provable.

Chip Manufacturing: One Company

TSMC's market share (TrendForce, Q2 2025):

  • Q1 2025: 67.6%
  • Q2 2025: 70.2%
  • Projected 2025: 66-70%

Taiwan Semiconductor Manufacturing Company controls seven out of every ten dollars spent in the foundry market. Q2 2025 revenue: $30.24 billion, up 18.5% quarter-over-quarter.

Second place? Samsung at 7.3%. Third? SMIC at 5.1%.

That's not a competitive market. That's one company with a chokehold on modern computing and everyone else fighting for table scraps.

GPUs: Nine Out of Ten

Jon Peddie Research data, 2025:

  • Q1: Nvidia 92%, AMD 8%, Intel 0%
  • Q2: Nvidia 94%, AMD 6%, Intel 0%
  • Q3: Nvidia 92%, AMD 7%, Intel 1%

Q2 alone: Nvidia sold 10.9 million discrete GPUs. AMD managed 0.7 million. Intel shipped so few they round to zero.

In any other industry, 92% market share triggers antitrust scrutiny. In semiconductors, it gets you a keynote slot and a leather jacket.

Memory: Three Companies, Seventy Percent

Samsung Electronics, SK hynix, and Micron Technology control approximately 70% of the global DRAM market (PC Gamer).

The only alternative worth mentioning, Chinese manufacturer CXMT, holds maybe 5-10% and faces US export controls. Everyone else is a rounding error.

Three companies. Seventy percent of all memory. And they all make the same strategic decisions at the same time, something I'll come back to, because it sure looks like coordination even when it technically isn't.

I've written about Samsung's role in this before. And the broader memory crunch. But those were individual threads. This is the whole tapestry.

The AI Demand Shock

This is where the oligopoly structure stops being an academic concern and starts costing you money.

Memory Reallocation

2024-2025: AI data centers drove hyperscalers to build servers at staggering rates, consuming massive amounts of High-Bandwidth Memory. HBM is more lucrative than commodity DRAM, so all three memory companies reallocated wafer capacity toward it. Simultaneously. Independently. Of course.

The result, per IntuitionLabs: DRAM prices tripled year-over-year by late 2025. Inventories shrank to 8 weeks of supply.

Each gigabyte of HBM consumes roughly three times the wafer capacity of DDR5 (Tom's Hardware, September 2025). Yield loss from stacking, smaller die sizes. Every HBM chip cannibalizes 3-4 consumer RAM chips.

AI doesn't just use more memory. It devours manufacturing capacity at a ratio that makes consumer DRAM a rounding error in the profit equation.

The Stargate Number

October 2025: OpenAI announced its "Stargate" AI infrastructure partnership with Samsung and SK hynix. The memory demand? Up to 40% of global DRAM output. Roughly 900,000 wafers per month.

One AI project. Forty percent of global DRAM production.

That's not demand. That's resource capture at industrial scale, the kind of allocation that reshapes entire supply chains in its wake. And your $420 RAM kit is the wake.

What Consumers Got

Prices:

  • Corsair Vengeance DDR5-6000 32GB: $134.99 in September → over $420 by early December
  • DRAM year-over-year: +172% by Q3 2025 (TrendForce)
  • Q1 2026 forecast: another 50-55% increase over Q4 2025

Analyst Tom Hsu told CNBC he'd never seen memory prices move this fast. I've been tracking this market through the initial price jumps, the pandemic shortages, and crypto mining. This eclipses all of them.

GPU cuts: Nvidia's slashing RTX 50-series production by 30-40% in H1 2026 due to GDDR7 shortages. The 5070 Ti and 5060 Ti 16GB face the sharpest cuts. Why? Memory suppliers would rather sell to data centers. I covered this in the GPU production cuts piece and the 2026 GPU drought analysis.

Micron exits consumers entirely: They're killing the Crucial brand by early 2026. Official rationale: "free up wafer supply for strategic accounts." Translation: consumers aren't strategic. You're not profitable enough to bother with.

A memory manufacturer posting record profits while abandoning the consumer market. That's the oligopoly thesis in one sentence.

Controlled Scarcity—or Just Oligopoly Economics?

I went in looking for smoking guns. What I found is that oligopoly economics don't need smoking guns. They produce the same outcomes through structure alone.

What's Documented

1. Concentration is extreme

  • TSMC: 70% of foundry market (TrendForce, June 2025)
  • Nvidia: 92% of discrete GPUs (Jon Peddie Research, 2025)
  • Samsung/SK hynix/Micron: ~70% of DRAM (PC Gamer)

Documented. Public. Not speculation.

2. AI triggers reallocation, not shortage

  • AI effectively consumes nearly 20% of global DRAM supply when factoring in HBM's wafer-equivalent usage (Commercial Times via TrendForce)
  • IDC calls it "a potentially permanent, strategic reallocation of the world's silicon wafer capacity"

Memory makers are choosing high-margin AI products over consumer goods. This is allocation, not shortage. The difference matters, one implies scarcity, the other implies a choice. They chose.

3. Pricing power follows concentration

Three companies, 70% of supply, demand outstripping capacity. Prices spike 172% in a year. Dell COO Jeff Clarke, January 2026: "We've never witnessed costs escalating at this pace." He was describing DRAM, hard drives, and NAND flash. All at once.

What's Not Proven

Explicit price-fixing: No evidence of coordinated manipulation has surfaced. Yet. The 2000s DRAM price-fixing cases proved it's happened in this exact industry before.

Artificial stockpiling: The "warehouse full of GPUs" conspiracy has no basis. Capacity constraints are real, ASML's EUV machines, TSMC's 2nm process limits, HBM packaging bottlenecks.

Hidden supply: Manufacturing bottlenecks are genuine. The conspiracy angle doesn't hold. The structural angle does.

The Gray Zone

Oligopolies don't need to collude. They just need to exist.

Samsung shifts to HBM. SK hynix shifts to HBM. Micron shifts to HBM. Same time. Same direction. Was that coordination? Legally, no, rational response to identical market signals. But the consumer outcome is indistinguishable from collusion: supply contracts, prices surge, and three companies post record profits while telling you it's the market.

When incentives align this perfectly, collusion is redundant. The structure does the work.

Why Nobody Enters

Because entry costs are nation-state level:

  • Modern fab: $10-20+ billion
  • One ASML EUV machine: ~$380 million, years-long waitlist
  • Time to production: 3-5 years minimum

Intel tried with their foundry business. Billions spent, still hemorrhaging market share. GlobalFoundries gave up on leading-edge entirely. The graveyard of would-be competitors is extensive and expensive.

The barrier isn't just capital, it's physics, talent, supply chains, and geopolitics compounding on each other. You can't brute-force your way into semiconductor manufacturing. This is one of the few industries where $20 billion might not clear the entry fee.

The Geopolitical Layer

Semiconductors aren't products anymore. They're strategic assets on par with energy reserves.

  • US-China export controls restricting advanced node access to Chinese fabs
  • CHIPS Act: $52 billion for domestic US capacity (I covered the EU's version, €43 billion spread thin enough to be aspirational rather than strategic)
  • Taiwan risk: 70% of foundry capacity sitting on an island China claims

The US accuses China of throttling rare earth exports. China controls the refining. Both sides are weaponizing supply chains while pretending this is still about free trade.

Where This Goes

2026: Memory prices keep climbing. Consumer GPU supply stays constrained. PC and phone costs rise, or OEMs quietly compress specs to hold price points (which is already happening).

2027-2028: New fabs come online, TSMC Arizona, Intel Ohio. Memory capacity expands. But AI demand grows faster. Pricing stabilizes at a permanently higher baseline. The era of cheap hardware is over.

Long term: Compute becomes "the new oil," a strategic resource controlled by a handful of companies, fought over by nation-states. Semiconductor sovereignty replaces free-market rhetoric. Capacity either catches up, or the AI bubble pops. Place your bets.

My Assessment

The conspiracy theorists are wrong about the mechanism but right about the outcome.

No shadowy cabal. No conference room of villains. What exists is worse: structural incentives that produce identical results without anyone needing to coordinate. TSMC controls 70% of foundry capacity, they don't need to artificially constrain supply. Natural bottlenecks do it for them. Three memory makers independently decide HBM margins beat DDR5 margins, and the effect on your wallet is indistinguishable from price-fixing.

What's infuriating is the framing. "Supply constrained" like it's an act of God. It's not. DRAM capacity exists, it's being allocated to AI accelerators instead of consumer PCs. That's a business decision dressed up as a natural disaster.

Micron is exiting the Crucial consumer brand while posting record profits. In plain language: "We're making so much money from AI that we don't want your business anymore." A memory manufacturer choosing which customers deserve to exist, that's not market dynamics, that's an eviction notice with a quarterly earnings call attached.

OpenAI's Stargate consuming 40% of global DRAM isn't market demand, it's resource capture. One company monopolizing scarce supply because they can outbid the entire consumer market. And with three manufacturers controlling 70% of production, there's zero competitive pressure to serve the customers getting priced out.

The inevitability framing is the real scam. "AI demand is high, prices must rise." Sure. But oligopolies also control capacity expansion. When raising prices on existing supply is more profitable than building new fabs, why build? That's the trap: under-investment disguised as constraint.

I've watched this pattern repeat. Crypto mining boom. Pandemic shortages. Same playbook every time—"extraordinary demand" while manufacturers quietly maximize margins instead of capacity. The excuse rotates. The outcome never does. $135 → $420 in three months while the companies responsible post record earnings.

The AI boom didn't create these oligopolies. It revealed how little resilience exists when concentrated markets face demand shocks, and what happens when the companies controlling scarce resources answer exclusively to whoever pays the most.

The data here isn't ambiguous. Three companies control 70% of memory production, one AI project consumes 40% of output, prices spike 172%, and a manufacturer abandons consumers while posting records. This isn't a market functioning under pressure. This is a market structured to extract maximum value from minimum competition, and AI gave it the demand shock to prove it.

What I Couldn't Verify

  • Whether memory manufacturers maintain informal communication channels beyond public signals. The 2000s DRAM price-fixing convictions prove it's happened before. No evidence of current activity, but absence of evidence isn't evidence of absence in an industry with documented cartel history.
  • TSMC's exact customer allocation between AI and consumer chips. They don't publish this. Industry estimates range wildly.
  • Whether the Stargate 40% DRAM consumption figure represents peak demand or phased rollout. The number comes from industry analysis, not OpenAI.
  • How much of the 172% price increase reflects genuine scarcity versus distributors and retailers stacking margins on top of manufacturer increases. Probably both.

Sources: TrendForce (TSMC market share, June 2025; DRAM pricing data), Jon Peddie Research (GPU market share, Q1-Q3 2025), Tom's Hardware (HBM wafer capacity analysis, September 2025), PC Gamer (DRAM market concentration), IntuitionLabs (RAM shortage analysis), CNBC (Tom Hsu interview, memory pricing), IDC (memory shortage crisis analysis, 2025), Dell earnings call (Jeff Clarke quote, January 2026), Commercial Times via TrendForce (AI DRAM consumption), The Register, BattleforgePC, PatentPC, Z2Data

If Samsung, SK hynix, Micron, TSMC, or Nvidia want to correct anything here, my inbox is open. I'll publish corrections with the same prominence as the original claims.