CITIC Securities: Computing Power Oversupply Concerns Are Unfounded; Domestic Computing Power Holds Differentiated Value
Miles Bennett
CITIC Securities argues in its July strategy note that the recent tech sell-off stems from deleveraging, not an AI reversal — computing-power oversupply fears are unfounded. Domestic chips serve as a differentiated 'Plan B,' and earnings season favors high-certainty plays.
Meta leasing compute — glut signal or structural upgrade?
The panic trigger was reports that Meta plans to lease spare compute. But CITIC Securities, citing SemiAnalysis, notes Meta signed over 5 GW of capacity in six months — roughly $250–300 billion in spending — with no sign of excess.
This means → Meta is monetizing legacy compute, not idling new capacity. Its superior cluster capability lets it earn 3–4× the revenue per megawatt versus peers.
In plain terms = inference demand is concentrating at the top — Anthropic and OpenAI. Meta selling compute is asset optimization, a structural shift, not a demand decline.
$700 billion in capex still climbing — where does the money come from?
In Q1 2026 the four major cloud providers raised full-year capex guidance again, totaling $700 billion, with an upbeat 2027 outlook.
With operating cash flow under pressure, the bond market has become the key funding channel. In H1 2026, five hyperscalers issued over $130 billion in debt. The Dallas Fed estimates full-year AI-related investment-grade issuance at roughly $300 billion.
This means → AI bonds compete with Treasuries for long-duration capital, potentially pushing up long-end yields and squeezing tech valuations. Financing won't cap capex growth itself, but the pace of issuance could weigh on tech multiples.
Are frontier models still improving? How far has China caught up?
CITIC Securities cites METR data showing Claude Mythos Preview now handles complex long-horizon tasks beyond the 16-hour test ceiling. The empirical rule — doubling task duration every seven months — has not slowed.
Claude Fable 5 and GPT-5.6 preview show clear gains in coding and agent capability over prior generations. On pricing, GPT-5.6 Sol outputs at $30 per million tokens, slightly below Fable 5's $50 per million tokens.
On the domestic side, GLM-5.2 is now the top Chinese open-source model, with coding ability in the global first tier. The note suggests markets may still underestimate Mythos / Fable 5 potential; AI4S, knowledge work, and physical AI could form a commercialization relay — but a "step-change moment" is needed for confirmation.
After the overseas AI crowd-trade, why would money flow to domestic compute?
CITIC Securities observes that offshore AI assets have entered a phase of high crowding, high correlation, and high volatility. International capital is searching for uncorrelated return sources.
This means → domestic compute, as a "Plan B," carries diversification value and could attract foreign allocation — it is not simply tracking offshore AI swings.
In plain terms = when offshore AI stocks move in lockstep, money naturally seeks an asset with a different rhythm. Domestic computing power fills that role.
Earnings season is here — which segments offer the highest certainty?
The note recommends focusing on segments with high earnings visibility and reasonable valuations: on a cyclical basis, domestic foundry (FAB) and semiconductor equipment, plus the lower-valued optical communications segment.
Along the price-hike chain, segments with high AI exposure that repriced first have stronger earnings delivery — watch memory and PCB upstream in particular.
This reflects a broader shift: as U.S. Q2 reports and A-share interim results arrive, pricing power returns to earnings delivery. The near-term pullback may offer entry points for leaders with earnings support; capex guidance from the four major cloud providers and Nvidia's data-center revenue will be the key benchmarks for global AI momentum.
Content is for reference only, not financial advice.