15.1 C
Delhi
Saturday, February 21, 2026

Are DeepSeek Moments Now the New Normal?

(Bloomberg Opinion) — A little-known Chinese AI company recently released an open-source reasoning model that challenged Western dominance and was developed at a fraction of the cost. And no, it’s not DeepSeek.

When Moonshot AI, a Beijing-based lab, launched Kimi K2 Thinking earlier this month, it went viral in tech circles. A partner at prominent Silicon Valley venture capitalist firm Menlo Ventures called it a “turning point in AI.” The model now ranks second on Artificial Analysis’ intelligence index, behind only OpenAI’s GPT 5.1 — and ahead of the latest offerings from Alibaba Group Holding Ltd. and DeepSeek, as well as US titans like X.AI and Anthropic. Looking at another benchmark measuring more complex, problem-solving “agentic” tasks, it even outperformed OpenAI.

This time around, however, markets barely batted an eye. As Bloomberg Economics Michael Deng noted, “The contrast with January’s DeepSeek panic, which wiped almost $600 billion off Nvidia in a single day, reveals how quickly investors have internalized that Chinese labs can match frontier capabilities at lower cost.” Have we already reached the point where matching the best in AI on a shoestring budget is no longer a shock?

It’s true that it has become increasingly hard to judge model performance based on benchmarks alone. Moonshot’s latest release joins an especially crowded domestic market. Launches and updates from Alibaba, Zhipu and MiniMax have come at a frenetic pace this year. The competition fuels innovation, even if it makes it difficult for one firm to stand out and gain a viable competitive edge — and the path to monetization seems elusive.

At the same time, the cost gap with the West is striking. Citing a source familiar with the matter, CNBC reported that Kimi K2 Thinking cost $4.6 million to train. A member of the Moonshot team later said in a Reddit Ask Me Anything session that this wasn’t “an official number.” But the representative did cheekily nod to the major spending differences in response to a question on when the next-generation model will be released, saying it will come “before Sam’s trillion-dollar data center is built.”

Silicon Valley has taken notice. I’ve written before that more US startups seem to be quietly building on Chinese AI models, including Moonshot’s. (Even before the latest update, venture capitalist Chamath Palihapitaya said that a company he worked with has switched over to Kimi K2.) It’s a trend that’s harder to quantify because few firms want to get caught up in the geopolitical crosshairs of US-China AI competition.

Fears of Communist Party censorship get a lot of attention, but testers have pointed out that this becomes less of an issue when you’re downloading and deploying the models locally. As a Moonshot representative put it in the Reddit AMA, “open-sourcing the model is hopefully a good step” to ease concerns about Chinese origins.

Part of the reason for their popularity is China’s low-cost, open-source approach. And while this allows developers to download and build on top of these models, doing so at scale will still require some amount of AI infrastructure. It means that for Nvidia Corp. and other chipmakers the threat isn’t as existential as some imagined after the DeepSeek-triggered selloff. This may partially explain some of the muted market reaction to the torrent of highly capable models from China.

But something still doesn’t quite add up looking at the valuation gulf. Despite its latest model’s performance coming very close to OpenAI’s, Moonshot’s most-recent valuation of some $3.3 billion is a rounding error next to the US titan’s $500 billion. Even the nine-month-old AI startup from former OpenAI executive Mira Murati is reportedly in funding talks pegging it at $50 billion. It makes increasing fears of an AI bubble seem valid.

Jefferies analysts noted last week that Chinese hyperscalers’ combined capital expenditures between 2023 and 2025 were 82% lower than US peers. But the performance gap between their two best models, based on various analysis, is now razor-thin. Even with inferior quality chips and high competition, the significantly lower spending points to a clearer path to return on investment emerging in China.

After the Kimi K2 update release, the co-founder of Hugging Face, Thomas Wolf, flagged its advancements in a social media post and asked: “Is this another DeepSeek moment?” He quickly followed that up with: “Should we expect this every couple [of] months now?”

The answer increasingly looks to be a resounding yes. We’ve normalized the idea that Chinese AI labs can seemingly come out of nowhere to close the performance gap with Silicon Valley despite chip constraints and significantly smaller budgets. For US tech giants, the question now isn’t whether they can stay ahead — it’s whether their massive spending outlays will actually translate into better commercial returns.

More From Bloomberg Opinion:

This column reflects the personal views of the author and does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.

Catherine Thorbecke is a Bloomberg Opinion columnist covering Asia tech. Previously she was a tech reporter at CNN and ABC News.

Latest

FBI warns of ATM Jackpotting incidents across America: What are they, how they work and how to detect one

Tech News News: The Federal Bureau of Investigation (FBI) has released a flash to disseminate indicators of compromise (IOCs) and technical details associated w

“Great meeting”: OpenAI CEO Sam Altman on meeting PM Modi, says “incredible energy around AI in India”

OpenAI CEO Sam Altman described his meeting with Prime Minister Narendra Modi on Friday as great and highlighted the remarkable momentum of artificial intellige

IT leaders discuss AIs impact on SaaS at India AI Impact Summit

IT leaders discuss AI's impact on SaaS at India AI Impact Summit

Many nations have lauded Indias move to mandate AI labelling, says Vaishnaw, as new IT rules take effect

Many nations have lauded India's move to mandate AI labelling, says Vaishnaw, as new IT rules take effect

US leads AI brain race followed by China, Singapore

US leads AI brain race followed by China, Singapore; India at 6th spot: Report

Topics

Trump to ditch IEEPA tariffs; uses alternative legal powers, says US Treasury Secretary Scott Bessent

After the US Supreme court strikes down Trump IEEPA Tariffs, President turns to Section 232, 301 & 122 for unchanged revenues.  

Trump signs 10% global tariffs order, calls Supreme Court ruling ‘deeply disappointing’

The new tariff will come into force almost immediately and will remain effective for approximately five months under Section 122 of the Trade Act of 1974.

Section 122, 301 and 232: Trump’s legal arsenal for new global tariffs

After the Supreme Court blocked his emergency tariffs, Trump moved to impose a 10% global duty under Section 122 and signaled broader trade action through Secti

Trump repeats India-Pak mediation claim while slamming tariff verdict

Trump claimed 200% tariff threats helped broker an India-Pakistan ceasefire, a claim India has firmly denied.

Which Trump tariffs did US Supreme Court strike down? Here’s the breakdown

The Supreme Court struck down Trump’s sweeping global tariffs imposed under emergency powers, voiding duties on major trading partners while leaving national

Michael Jackson in Epstein files? Truth behind claim of Neverland protecting kids from Epstein Island; ‘he was silenced’

Many are wondering if Michael Jackson is in the Epstein files amid claims that he used Neverland ranch to protect kids from what happened on Epstein Island.

Americans could save $900 this year after Supreme Court’s tariff ruling, experts weigh in

The SCOTUS blocked key tariffs imposed by Donald Trump, potentially lowering household costs

Tariff rebate check update: Big warning on $2000 payment amid Trump vs SCOTUS; ‘got a feeling’

President Donald Trump’s Treasury Secretary Scott Bessent appeared to give his verdict on the $2000 rebate checks
spot_img

Related Articles

Popular Categories

spot_imgspot_img