Chinese AI agency DeepSeeok has emerged as a possible challenger to U.S. AI leaders, demonstrating breakthrough fashions that declare to supply efficiency akin to main chatbots at a fraction of the associated fee. The firm’s cellular app, launched in early January, has additionally topped iPhone charts throughout main markets together with the U.S., UK, and China.
Founded in 2023 by Liang Wenfeng, former chief of AI-driven quant hedge fund High-Flyer, DeepSeeok makes its fashions open-source and incorporates a reasoning characteristic that articulates its considering earlier than offering responses.
Wall Street’s response has been blended. While Jefferies warns that DeepSeeok’s environment friendly strategy “punctures among the capex euphoria” following current spending commitments from Meta and Microsoft — every exceeding $60 billion this yr — Citi questions whether or not such outcomes have been achieved with out superior GPUs. Goldman Sachs sees broader implications, suggesting the event may reshape competitors between established tech giants and startups by reducing obstacles to entry.
Here’s how Wall Street analysts are reacting to DeepSeeok, in their very own phrases (emphasis mine):
Jefferies
DeepSeeok’s energy implications for AI coaching punctures among the capex euphoria which adopted main commitments from Stargate and Meta final week. With DeepSeeok delivering efficiency akin to GPT-4o for a fraction of the computing energy, there are potential destructive implications for the builders, as stress on AI gamers to justify ever rising capex plans may finally result in a decrease trajectory for knowledge middle income and revenue progress.
If smaller fashions can work effectively, it’s doubtlessly optimistic for smartphone. We are bearish on AI smartphone as AI has gained no traction with shoppers. More {hardware} improve (adv pkg+quick DRAM) is required to run greater fashions on the telephone, which is able to increase prices. AAPL’s mannequin is actually based mostly on MoE, however 3bn knowledge parameters are nonetheless too small to make the companies helpful to shoppers. Hence DeepSeeok’s success presents some hope however there isn’t any influence on AI smartphone’s near-term outlook.
China is the solely market that pursues LLM effectivity owing to chip constraint. Trump/Musk probably acknowledge the danger of additional restrictions is to power China to innovate quicker. Therefore, we predict it probably Trump will calm down the AI Diffusion coverage.
Citi
While DeepSeeok’s achievement could possibly be groundbreaking, we query the notion that its feats have been carried out with out the usage of superior GPUs to advantageous tune it and/or construct the underlying LLMs the ultimate mannequin relies on via the Distillation method. While the dominance of the US corporations on essentially the most superior AI fashions could possibly be doubtlessly challenged, that stated, we estimate that in an inevitably extra restrictive surroundings, US’ entry to extra superior chips is a bonus. Thus, we don’t anticipate main AI corporations would transfer away from extra superior GPUs which offer extra engaging $/TFLOPs at scale. We see the current AI capex bulletins like Stargate as a nod to the necessity for superior chips.
Bernstein
In brief, we consider that 1) DeepSeeok DID NOT “construct OpenAI for $5M”; 2) the fashions look unbelievable however we don’t assume they’re miracles; and three) the ensuing Twitterverse panic over the weekend appears overblown.
Our personal preliminary response doesn’t embrace panic (removed from it). If we acknowledge that DeepSeeok might have lowered prices of reaching equal mannequin efficiency by, say, 10x, we additionally be aware that present mannequin price trajectories are rising by about that a lot yearly anyway (the notorious “scaling legal guidelines…”) which might’t proceed perpetually. In that context, we NEED improvements like this (MoE, distillation, blended precision and many others) if AI is to proceed progressing. And for these searching for AI adoption, as semi analysts we’re agency believers within the Jevons paradox (i.e. that effectivity beneficial properties generate a web improve in demand), and consider any new compute capability unlocked is way extra prone to get absorbed attributable to utilization and demand improve vs impacting long run spending outlook at this level, as we don’t consider compute wants are anyplace near reaching their restrict in AI. It additionally looks like a stretch to assume the improvements being deployed by DeepSeeok are utterly unknown by the huge variety of prime tier AI researchers on the world’s different quite a few AI labs (frankly we don’t know what the big closed labs have been utilizing to develop and deploy their very own fashions, however we simply can’t consider that they haven’t thought of and even maybe used related methods themselves).
Morgan Stanley
We haven’t confirmed the veracity of those stories, but when they’re correct, and superior LLM are certainly in a position to be developed for a fraction of earlier funding, we may see generative AI run finally on smaller and smaller computer systems (downsizing from supercomputers to workstations, workplace computer systems, and at last private computer systems) and the SPE business may benefit from the accompanying improve in demand for associated merchandise (chips and SPE) as demand for generative AI spreads.
Goldman Sachs
With the newest developments, we additionally see 1) potential competitors between capital-rich web giants vs. start-ups, given reducing obstacles to entry, particularly with current new fashions developed at a fraction of the price of present ones; 2) from coaching to extra inferencing, with elevated emphasis on post-training (together with reasoning capabilities and reinforcement capabilities) that requires considerably decrease computational sources vs. pre-training; and three) the potential for additional world enlargement for Chinese gamers, given their efficiency and value/value competitiveness.
We proceed to anticipate the race for AI utility/AI brokers to proceed in China, particularly amongst To-C functions, the place China corporations have been pioneers in cellular functions within the web period, e.g., Tencent’s creation of the Weixin (WeChat) super-app. Amongst To-C functions, ByteDance has been main the way in which by launching 32 AI functions over the previous yr. Amongst them, Doubao has been the most well-liked AI Chatbot to date in China with the very best MAU (c.70mn), which has not too long ago been upgraded with its Doubao 1.5 Pro mannequin. We consider incremental income streams (subscription, promoting) and eventual/sustainable path to monetization/optimistic unit economics amongst functions/brokers can be key.
For the infrastructure layer, investor focus has centered round whether or not there can be a near-term mismatch between market expectations on AI capex and computing demand, within the occasion of serious enhancements in price/mannequin computing efficiencies. For Chinese cloud/knowledge middle gamers, we proceed to consider the main target for 2025 will focus on chip availability and the flexibility of CSP (cloud service suppliers) to ship enhancing income contribution from AI-driven cloud income progress, and past infrastructure/GPU renting, how AI workloads & AI associated companies may contribute to progress and margins going ahead. We stay optimistic on long-term AI computing demand progress as an additional reducing of computing/coaching/inference prices may drive larger AI adoption. See additionally Theme #5 of our key themes report for our base/bear situations for BBAT capex estimates relying on chip availability, the place we anticipate mixture capex progress of BBAT to proceed in 2025E in our base case (GSe: +38% yoy) albeit at a barely extra average tempo vs. a powerful 2024 (GSe: +61% yoy), pushed by ongoing funding into AI infrastructure.
J.P.Morgan
Above all, a lot is made from DeepSeeok’s analysis papers, and of their fashions’ effectivity. It’s unclear to what extent DeepSeeok is leveraging High-Flyer’s ~50k hopper GPUs (related in dimension to the cluster on which OpenAI is believed to be coaching GPT-5), however what appears probably is that they’re dramatically decreasing prices (inference prices for his or her V2 mannequin, for instance, are claimed to be 1/7 that of GPT-4 Turbo). Their subversive (although not new) declare – that began to hit the US AI names this week – is that “extra investments don’t equal extra innovation.” Liang: “Right now I don’t see any new approaches, however large corporations shouldn’t have a transparent higher hand. Big corporations have present prospects, however their cash-flow companies are additionally their burden, and this makes them weak to disruption at any time.” And when requested about the truth that GPT5 has nonetheless not been launched: “OpenAI is just not a god, they received’t essentially all the time be on the forefront.”
UBS
Throughout 2024, the primary yr we noticed huge AI coaching workload in China, greater than 80-90% IDC demand was pushed by AI coaching and concentrated in 1-2 hyperscaler prospects, which translated to wholesale hyperscale IDC demand in comparatively distant space (as power-consuming AI coaching is delicate to utility price quite than consumer latency).
If AI coaching and inference price is considerably decrease, we might anticipate extra finish customers would leverage AI to enhance their enterprise or develop new use circumstances, particularly retail prospects. Such IDC demand means extra concentrate on location (as consumer latency is extra vital than utility price), and thus larger pricing energy for IDC operators which have ample sources in tier 1 and satellite tv for pc cities. Meanwhile, a extra diversified buyer portfolio would additionally indicate larger pricing energy.
We’ll replace the story as extra analysts react.
