CITIC Securities: AI Stock Correction Stems from Forward Pricing, Three Key Narrative Variables Await Clarification
Nashnova编辑部
CITIC Securities argues the tech selloff is not about high long-bond yields but about shaken confidence in AI stocks' forward pricing — commercialization pace, market-share durability, and the model-generation gap remain unresolved, with anti-distillation as the single biggest wildcard.
High bond yields aren't the real culprit — so what is?
The Magnificent Seven have issued $219.2 billion in debt so far in 2026 — up 150.6% from all of 2025. This means → big tech itself is the force pushing long-end yields higher. Blaming those same yields for the stock decline is circular logic.
In plain terms = rising long-bond yields and a tech-company borrowing spree are two sides of the same AI capex boom, not cause and effect.
CITIC Securities sees the real drag as the market questioning whether AI stocks can deliver on their forward valuations — not the cost of funding itself.
Narrative variable one: is AI actually making money?
Anthropic's ARR — annualized recurring revenue, the current payment run-rate projected over a year — grew roughly 18% month-on-month from May to July, a slight slowdown. But OpenAI CFO Sarah Friar said quarter-to-date ARR growth is about 35%. This means → taken together, overall commercialization momentum has not materially slowed.
A chunk of revenue is invisible: token consumption sold through cloud providers (TaaS — Token as a Service) does not appear in model makers' own ARR. In plain terms = looking only at model makers' books systematically understates real end-user spending growth.
CITIC Securities warns, however, that until AI agents unlock compelling paid use cases beyond coding, the TaaS narrative alone will struggle to pull in fresh capital.
Narrative variable two: can compute dominance lock in market share?
According to Ramp's AI Index — tracking API spend across 70,000+ U.S. companies — OpenAI's share rose from 28.5% in May to 36.0% in July. Nearly all the gain came from GPT-5.6 Sol, which went from zero share in May to 14.9% alone in July.
This reflects how quickly a single strong model can reshape the competitive map. But the reverse is equally true: if frontier models converge in capability and switching costs stay low, share can be rewritten just as fast.
In plain terms = compute advantage delivers episodic share gains, not a moat — the question ultimately comes back to how big the total pie can grow.
Narrative variable three: anti-distillation — why is it the biggest wildcard?
In mid-August 2026, researchers from MATS Research, the ELLIS Tübingen Institute, and other groups published a paper analyzing how extractable frontier-model reasoning chains — the step-by-step intermediate logic a model uses to reach an answer — are under current mainstream API architectures. This means → the reasoning-capability moat that big labs built with massive compute faces the risk of low-cost replication.
If the anti-distillation problem remains unsolved, the pricing logic for compute "pick-and-shovel" players (companies selling AI infrastructure, such as Nvidia) degrades toward a traditional public-utility model — margins compress.
If frontier labs solve anti-distillation by year-end and ship a clearly stronger next-generation model, training-side scaling advantages convert into a durable competitive barrier and the compute arms race intensifies further. CITIC Securities calls this the single most important pricing variable.
U.S. Treasury buybacks and A-share positioning — what else to watch?
The U.S. Treasury raised its long-bond buyback size from $2 billion to $4 billion. But publicly held Treasury debt now exceeds $32 trillion — $4 billion is a drop in the bucket. This means → the move is more signal than substance: it reinforces the expectation that Treasury will intervene when long-end rates spiral, yet may also deepen doubts about fiscal discipline.
On the A-share side, active hedge funds added 7.3 percentage points of equity exposure in the first week of August. Large funds' equity-position index hit 88.56%, with 77.11% fully invested — both year-to-date highs. In plain terms = the most aggressive money is already in and drove this rally; bullish expectations are largely priced in.
CITIC Securities recommends managing expectations in a range-bound market. Within tech, rotate from AI pricing-uplift plays toward "volume certainty" assets — gas turbines, wafer foundries, semiconductor equipment. Outside tech, add energy-chemicals, non-ferrous metals, innovative pharma, and leading brokerages with overseas expansion potential.
Content is for reference only, not financial advice.