Goldman Sachs: Kimi K3 Confirms the End of the Scaling Era
N.R. Finch
Goldman partner Rich Privorotsky calls Moonshot AI's Kimi K3 a paradigm shock on par with last year's DeepSeek moment — scaling is no longer the only game, and the rules of AI competition are being rewritten.
What makes Kimi K3 so significant?
Kimi K3 leads Anthropic's flagship Opus 4.8 on Arena's broader text rankings, yet is priced 40% lower.
Moonshot AI plans to release it on July 27 with open-source weights. This means → enterprises and governments can run and customize the model on their own servers, with no reliance on U.S. cloud providers.
In plain terms = a model that matches the frontier in performance, costs 40% less, and can be self-hosted is already "good enough and cheaper" for most buyers.
Why is the "scaling era" ending?
Privorotsky is clear: scaling laws — the principle that bigger models plus more data yield better performance — haven't failed mathematically. But the marginal return on each extra pre-training dollar is falling.
The real capability gains are shifting to reinforcement learning (RL) and post-training — the fine-tuning stage after a model's initial training run.
This reflects what OpenAI co-founder Ilya Sutskever asked nine months ago: is AI moving from a "scaling era" back to a "research era"?
What does CXMT's IPO mean?
Chinese memory-chip maker CXMT (长鑫存储) is listing on Shanghai's STAR Market this month, with its IPO oversubscribed more than 200 times.
This means → China is building a fourth scaled DRAM competitor after Samsung, SK Hynix, and Micron.
More capital, more capacity, and another aggressive entrant — investors must now reassess how long the memory-supply bottleneck can last and whether incumbents' margins will be squeezed.
What drove the chip-stock selloff?
Chip stocks fell sharply overnight. Privorotsky characterizes it as a "de-grossing event" — funds cutting leveraged positions across the board.
In plain terms = fundamentals didn't suddenly deteriorate; leverage amplified the selling, and the most crowded trades led the decline.
The next checkpoint comes after options expiry: will a drop in dealer gamma exposure let the hardware de-leveraging spill into the broader market, or will unwinding hedges create room for a bounce?
Are hyperscalers really the winners?
Privorotsky says he is puzzled by the rush into hyperscalers — mega cloud platforms such as Amazon AWS and Microsoft Azure.
The bull case: lower compute costs → cloud providers deliver the same AI capability with less capital → hardware de-rating is just the market pricing in cheaper forward compute, with value migrating to the platform layer.
But his reservation matters: inference is becoming a fiercely competitive business. As models commoditize and API pricing compresses, "if your product is just tokens, holding margins gets harder and harder."
What is the core takeaway for investors?
Even if U.S. labs pull ahead again, China has proved it can close the gap quickly — Kimi K3 and DeepSeek confirm this back to back.
Scaling laws may still hold mathematically, but the economic question is whether pouring 10× or even 100× more pre-training compute can generate incremental capability worth the cost.
This signals something deeper: AI's core moat is shifting from "who can spend the most" to "whose research is the most efficient."
Content is for reference only, not financial advice.