AI Chip Production Outpaces Data Center Capacity, Pressuring Semiconductor Shipment Expectations
nashnova research
Morgan Stanley warns that chip production already exceeds available data-center space, with the U.S. power shortfall projected to reach roughly one-third of the electricity needed for chip sales by 2028 — raising the risk that bullish semiconductor shipment forecasts are proven wrong, with knock-on effects across chipmakers, data-center developers, and cloud tenants.
Chips are rolling off the line — but where do they plug in?
Morgan Stanley's core call: chip production is now outrunning data-center construction, creating a supply-demand mismatch.
This means → the problem is not demand for chips — it is that buyers' "buildings" are not ready. The hardware arrives with nowhere to power up.
The U.S. data-center power gap is projected to hit roughly one-third of the electricity chip sales would require by 2028.
In plain terms = even if every chip sells, a third of the compute cannot run because there is no electricity.
Why are data centers so far behind schedule?
According to Currence, half of the large data-center projects announced this year have not broken ground.
The obstacle list is long: equipment and labor shortages + community opposition + construction bans + permitting delays + slow grid connections + gas-turbine capacity constraints.
BloombergNEF data adds another worry: nearly half of the U.S. data-center pipeline under construction is led by developers entering the sector for the first time.
This reflects an expansion pace that has outstripped experienced operators — a high share of first-timers amplifies delay risk further.
Who do contracts actually protect?
Data-center leases typically limit tenants' exit rights and share cost overruns, but the strength of protection varies contract by contract.
Nscale signed a $44.6 billion compute contract with Anthropic, yet the terms state that supply-chain delays offer "limited remedy" — Anthropic can terminate "without liability."
This means → for the developer, revenue is uncertain while the client can walk at any time; for the tenant, the delivery date is not guaranteed.
Oracle's Jupiter project in New Mexico was delayed by gas-pipeline permitting; Oracle issued a force-majeure notice but still owes partial payments to project lenders even if rent is deferred.
SoftBank's SB Energy — $50 billion valuation, zero operating data centers?
SB Energy's IPO prospectus discloses a contracted pipeline of 8.8 GW, mostly serving OpenAI — yet only a tiny fraction is under construction and no data center is operational.
This means → big pipeline, high valuation, near-zero delivery capacity — the $50 billion valuation is being questioned by investors.
Nvidia stepped in with a $105 billion guarantee, backing OpenAI's lease obligations to help SB Energy move forward.
This reflects Nvidia's own anxiety: if data centers cannot be built, its processors ultimately cannot sell. Nvidia's latest filing lists shortages in land, power, construction, or capital as risk factors.
The hyperscalers' "stockpiling" dilemma
Amazon, Microsoft, and peers typically buy servers only months before a data center goes live — they do not want to hoard chips that cannot be powered on and risk obsolescence.
AI companies now hold nearly $350 billion in assets not yet in use.
In plain terms = equipment bought with real money is sitting in warehouses, generating no revenue and losing value.
BloombergNEF suggests swapping older chips for Nvidia's latest, more efficient models to bridge the capacity gap — but that would push capex and depreciation even higher. Another option: shifting data-center capacity to countries with more abundant renewable energy.
Nvidia near $6 trillion — what is the market betting on?
Nvidia's market cap is approaching $6 trillion; investors do not appear overly worried about its ability to find buyers for high-margin processors.
The real test: whether power and construction bottlenecks can be cleared before chip-shipment momentum stalls.
This means → the market's current optimistic pricing is not a bet that chip demand does not exist — it is a bet that infrastructure can keep up. That assumption faces a growing body of contrary evidence.
市场有风险，内容仅供研究参考，不构成投资建议。