DeepSeek
China’s open source AI models to capture a larger share of 2026 global AI market
Overview of AI Models – China vs U.S. :
Chinese AI language models (LMs) have advanced rapidly and are now contesting with the U.S. for global market leadership. Alibaba’s Qwen-Image-2512 is emerging as a top-performing, free, open-source model capable of high-fidelity human, landscape, and text rendering. Other key, competitive models include Zhipu AI’s GLM-Image (trained on domestic chips), ByteDance’s Seedream 4.0, and UNIMO-G.
Today, Alibaba-backed Moonshot AI released an upgrade of its flagship AI model, heating up a domestic arms race ahead of an expected rollout by Chinese AI hotshot DeepSeek. The latest iteration of Moonshot’s Kimi can process text, images, and videos simultaneously from a single prompt, the company said in a statement, aligning with a trend toward so-called omni models pioneered by industry leaders like OpenAI and Alphabet Inc.’s Google.
Moonshot AI Kimi website. Photographer: Raul Ariano/Bloomberg
………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………..
Chinese AI models are rapidly narrowing the gap with Western counterparts in quality and accessibility. That shift is forcing U.S. AI leaders like Alphabet’s Google, Microsoft’s Copilot, OpenAI, and Anthropic to fight harder to maintain their technological lead in AI. That’s despite their humongous spending on AI data centers, related AI models and infrastructure.
In early 2025, investors seized on DeepSeek’s purportedly lean $5.6 million LM training bill as a sign that Nvidia’s high-end GPUs were already a relic and that U.S. hyperscalers had overspent on AI infrastructure. Instead, the opposite dynamic played out: as models became more capable and more efficient, usage exploded, proving out a classic Jevons’ Paradox and validating the massive data-center build-outs by Microsoft, Amazon, and Google.
The real competitive threat from DeepSeek and its peers is now coming from a different direction. Many Chinese foundation models are released as “open source” or “open weight” AI models which makes them effectively free to download, easy to modify, and cheap to run at scale. By contrast, most leading U.S. players keep tight control over their systems, restricting access to paid APIs and higher-priced subscriptions that protect margins but limit diffusion.
That strategic divergence is visible in how these systems are actually used. U.S. models such as Google’s Gemini, Anthropic’s Claude, and OpenAI’s GPT series still dominate frontier benchmarks [1′] and high‑stakes reasoning tasks. According to a recently published report by OpenRouter, a third-party AI model aggregator, and venture capital firm Andreessen Horowitz. Chinese open-source models have captured roughly 30% of the “working” AI market. They are especially strong in coding support and roleplay-style assistants—where developers and enterprises optimize for cost efficiency, local customization, and deployment freedom rather than raw leaderboard scores.
Note 1. A frontier benchmark for AI models is a high-difficulty evaluation designed to test the absolute limits of artificial intelligence in complex,, often unsolved, reasoning tasks. FrontierMath, for example, is a prominent benchmark focusing on expert-level mathematics, requiring AI to solve hundreds of unpublished problems that challenge, rather than merely measure, current capabilities.
China’s open playbook:
China’s more permissive stance on model weights is not just a pricing strategy — it’s an acceleration strategy. Opening weights turns the broader developer community into an extension of the R&D pipeline, allowing users to inspect internals, pressure‑test safety, and push incremental improvements upstream.
As Kyle Miller at Georgetown’s Center for Security and Emerging Technology argues, China is effectively trading away some proprietary control to gain speed and breadth: by letting capability diffuse across the ecosystem, it can partially offset the difficulty of going head‑to‑head with tightly controlled U.S. champions like OpenAI and Anthropic. That diffusion logic is particularly potent in a system where state planners, big tech platforms, and startups are all incentivized to show visible progress in AI.
Even so, the performance gap has not vanished. Estimates compiled by Epoch AI suggest that Chinese models, on average, trail leading U.S. releases by about seven months. The window briefly narrowed during DeepSeek’s R1 launch in early 2025, when it looked like Chinese labs might have structurally compressed the lag; since then, the gap has widened again as U.S. firms have pushed ahead at the frontier.
Capital, chips, and the power problem:
The reason the U.S. lead has held is massive AI infrastructure spending. Consensus forecasts put capital expenditure by largely American hyperscalers at roughly $400 billion in 2025 and more than $520 billion in 2026, according to Goldman Sachs Research. By comparison, UBS analysts estimate that China’s major internet platforms collectively spent only about $57 billion last year—a fraction of U.S. outlays, even if headline Chinese policy rhetoric around AI is more aggressive.
But sustaining that level of investment runs into a physical constraint that can’t be hand‑waved away: electricity. The newest data-center designs draw more than a gigawatt of power each—about the output of a nuclear reactor—turning grid capacity into a strategic bottleneck. China now generates more than twice as much power as the U.S., and its centralized planning system can more readily steer incremental capacity toward AI clusters than America’s fragmented, heavily regulated electricity market.
That asymmetry is already shaping how some on Wall Street frame the race. Christopher Woods, global head of equity strategy at Jefferies, recently reiterated that China’s combination of open‑source models and abundant cheap power makes it a structurally formidable AI competitor. In his view, the “DeepSeek moment” of early last year remains a warning that markets have largely chosen to ignore as they rotate back into U.S. AI mega‑caps.
A fragile U.S. AI advantage:
For now, U.S. companies still control the most important chokepoint in the stack: advanced AI accelerators. Access to Nvidia’s cutting‑edge GPUs remains a decisive advantage. Yesterday, Microsoft announced the Maia 200 chip – their first silicon and system platform optimized specifically for AI inference. The chip was was designed for efficiency, both in terms of its ability to deliver tokens per dollar and performance per watt of power used.
“Maia 200 can deliver 30% better performance per dollar than the latest generation hardware in our fleet today,” Microsoft EVP for Cloud and AI Scott Guthrie wrote in a blog post.
Image Credit: Microsoft
……………………………………………………………………………………………………………………………………………………………………………………………………………………………….
Leading Chinese AI research labs have struggled to match training results using only domestic designed silicon. DeepSeek, which is developing the successor to its flagship model and is widely expected to release it around Lunar New Year, reportedly experimented with chips from Huawei and other local vendors before concluding that performance was inadequate and turning to Nvidia GPUs for at least part of the training run.
That reliance underscores the limits of China’s current self‑reliance push—but it also shouldn’t be comforting to U.S. strategists. Chinese firms are actively working around the hardware gap, not waiting for it to close. DeepSeek’s latest research focuses on training larger models with fewer chips through more efficient memory design, an incremental but important reminder that architectural innovation can partially offset disadvantages in raw compute.
From a technology‑editorial perspective, the underlying story is not simply “China versus the U.S.” at the model frontier. It is a clash between two AI industrial strategies: an American approach that concentrates capital, compute, and control in a handful of tightly integrated platforms, and a Chinese approach that leans on open weights, diffusion, and state‑backed infrastructure to pull the broader ecosystem forward.
The question for 2026 is whether U.S. AI firms’ lead in capability and chips can keep outrunning China’s advantages in openness and power—or whether the market will again wait for a shock like DeepSeek to re‑price that risk.
Deepseek and Other Chinese AI Models:
DeepSeek published research this month outlining a method of training larger models using fewer chips through a more efficient memory design. “We view DeepSeek’s architecture as a new, promising engineering solution that could enable continued model scaling without a proportional increase in GPU capacity,” wrote UBS analyst Timothy Arcuri.
Export controls haven’t prevented Chinese companies from training advanced models, but challenges emerge when the models are deployed at scale. Zhipu AI, which released its open-weight GLM 4.7 model in December, said this month it was rationing sales of its coding product to 20% of previous capacity after demand from users overwhelmed its servers.
Moonshot, Zhipu AI and MiniMax Group Inc are among a handful of AI LM front-runners in a hotly contested battle among Chinese large language model makers, which at one point was dubbed the “War of One Hundred Models.”
“I don’t see compute constraints limiting [Chinese companies’] ability to make models that are better and compete near the U.S. frontier,” Georgetown’s Miller says. “I would say compute constraints hit on the wider ecosystem level when it comes to deployment.”
Gaining access to Nvidia AI chips:
U.S. President Donald Trump’s plan to allow Nvidia to sell its H200 chips to China could be pivotal. Alibaba Group and ByteDance, TikTok’s parent company, have privately indicated interest in ordering more than 200,000 units each, Bloomberg reported. The H200 outperforms any Chinese-produced AI chip, with a roughly 32% processing-power advantage over Huawei’s Ascend 910C.
With access to Nvidia AI chips, Chinese labs could build AI-training supercomputers as capable as American ones at 50% extra cost compared with U.S.-made ones, according to the Institute for Progress. Subsidies by the Chinese government could cover that differential, leveling the playing field, the institute says.
Conclusions:
A combination of open-source innovation and loosened chip controls could create a cheaper, more capable Chinese AI ecosystem. The possibility is emerging just as OpenAI and Anthropic consider public stock listings (IPOs) and U.S. hyperscalers such as Microsoft and Meta Platforms face pressure to justify heavy spending.
The risk for U.S. AI leaders is no longer theoretical; China’s open‑weight, low‑cost model ecosystem is already eroding the moat that Google, OpenAI, and Anthropic thought they were building. By prioritizing diffusion over tight control, Chinese firms are seeding a broad developer base, compressing iteration cycles, and normalizing expectations that powerful models should be cheap—or effectively free—to adapt and deploy.
U.S. AI leaders could face pressure on pricing and profit margins from China AI competitors while having to deal with AI infrastructure costs and power constraints. Their AI advantage remains real, but fragile—highly exposed to regulatory shifts, export controls, and any breakthrough in China’s workarounds on hardware and training efficiency. The uncomfortable prospect for U.S. AI incumbents is that they could win the race for the best models and still lose ground in the market if China’s diffusion‑driven strategy defines how AI is actually consumed at scale.
…………………………………………………………………………………………………………………………………………………………………………………………………………………..
References:
https://www.barrons.com/articles/deepseek-ai-gemini-chatgpt-stocks-ccde892c
https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference/
China gaining on U.S. in AI technology arms race- silicon, models and research
U.S. export controls on Nvidia H20 AI chips enables Huawei’s 910C GPU to be favored by AI tech giants in China
Goldman Sachs: Big 3 China telecom operators are the biggest beneficiaries of China’s AI boom via DeepSeek models; China Mobile’s ‘AI+NETWORK’ strategy
Bloomberg: China Lures Billionaires Into Race to Catch U.S. in AI
Comparing AI Native mode in 6G (IMT 2030) vs AI Overlay/Add-On status in 5G (IMT 2020)
Sources: AI is Getting Smarter, but Hallucinations Are Getting Worse
Recent reports suggest that AI hallucinations—instances where AI generates false or misleading information—are becoming more frequent and present growing challenges for businesses and consumers alike who rely on these technologies. More than two years after the arrival of ChatGPT, tech companies, office workers and everyday consumers are using A.I. bots for an increasingly wide array of tasks. But there is still no way of ensuring that these systems produce accurate information.
A groundbreaking study featured in the PHARE (Pervasive Hallucination Assessment in Robust Evaluation) dataset has revealed that AI hallucinations are not only persistent but potentially increasing in frequency across leading language models. The research, published on Hugging Face, evaluated multiple large language models (LLMs) including GPT-4, Claude, and Llama models across various knowledge domains.
“We’re seeing a concerning trend where even as these models advance in capability, their propensity to hallucinate remains stubbornly present,” notes the PHARE analysis. The comprehensive benchmark tested models across 37 knowledge categories, revealing that hallucination rates varied significantly by domain, with some models demonstrating hallucination rates exceeding 30% in specialized fields.
Hallucinations are when AI bots produce fabricated information and present it as fact. Photo Credit: More SOPA Images/LightRocket via Getty Images
Today’s A.I. bots are based on complex mathematical systems that learn their skills by analyzing enormous amounts of digital data. These systems use mathematical probabilities to guess the best response, not a strict set of rules defined by human engineers. So they make a certain number of mistakes. “Despite our best efforts, they will always hallucinate,” said Amr Awadallah, the chief executive of Vectara, a start-up that builds A.I. tools for businesses, and a former Google executive.
“That will never go away,” he said. These AI bots do not — and cannot — decide what is true and what is false. Sometimes, they just make stuff up, a phenomenon some A.I. researchers call hallucinations. On one test, the hallucination rates of newer A.I. systems were as high as 79%.
Amr Awadallah, the chief executive of Vectara, which builds A.I. tools for businesses, believes A.I. “hallucinations” will persist.Credit…Photo credit: Cayce Clifford for The New York Times
AI companies like OpenAI, Google, and DeepSeek have introduced reasoning models designed to improve logical thinking, but these models have shown higher hallucination rates compared to previous versions. For more than two years, those companies steadily improved their A.I. systems and reduced the frequency of these errors. But with the use of new reasoning systems, errors are rising. The latest OpenAI systems hallucinate at a higher rate than the company’s previous system, according to the company’s own tests.
For example, OpenAI’s latest models (o3 and o4-mini) have hallucination rates ranging from 33% to 79%, depending on the type of question asked. This is significantly higher than earlier models, which had lower error rates. Experts are still investigating why this is happening. Some believe that the complex reasoning processes in newer AI models may introduce more opportunities for errors.
Others suggest that the way these models are trained might be amplifying inaccuracies. For several years, this phenomenon has raised concerns about the reliability of these systems. Though they are useful in some situations — like writing term papers, summarizing office documents and generating computer code — their mistakes can cause problems. Despite efforts to reduce hallucinations, AI researchers acknowledge that hallucinations may never fully disappear. This raises concerns for applications where accuracy is critical, such as legal, medical, and customer service AI systems.
The A.I. bots tied to search engines like Google and Bing sometimes generate search results that are laughably wrong. If you ask them for a good marathon on the West Coast, they might suggest a race in Philadelphia. If they tell you the number of households in Illinois, they might cite a source that does not include that information. Those hallucinations may not be a big problem for many people, but it is a serious issue for anyone using the technology with court documents, medical information or sensitive business data.
“You spend a lot of time trying to figure out which responses are factual and which aren’t,” said Pratik Verma, co-founder and chief executive of Okahu, a company that helps businesses navigate the hallucination problem. “Not dealing with these errors properly basically eliminates the value of A.I. systems, which are supposed to automate tasks for you.”
For more than two years, companies like OpenAI and Google steadily improved their A.I. systems and reduced the frequency of these errors. But with the use of new reasoning systems, errors are rising. The latest OpenAI systems hallucinate at a higher rate than the company’s previous system, according to the company’s own tests.
The company found that o3 — its most powerful system — hallucinated 33% of the time when running its PersonQA benchmark test, which involves answering questions about public figures. That is more than twice the hallucination rate of OpenAI’s previous reasoning system, called o1. The new o4-mini hallucinated at an even higher rate: 48 percent.
When running another test called SimpleQA, which asks more general questions, the hallucination rates for o3 and o4-mini were 51% and 79%. The previous system, o1, hallucinated 44% of the time.
In a paper detailing the tests, OpenAI said more research was needed to understand the cause of these results. Because A.I. systems learn from more data than people can wrap their heads around, technologists struggle to determine why they behave in the ways they do.
“Hallucinations are not inherently more prevalent in reasoning models, though we are actively working to reduce the higher rates of hallucination we saw in o3 and o4-mini,” a company spokeswoman, Gaby Raila, said. “We’ll continue our research on hallucinations across all models to improve accuracy and reliability.”
Tests by independent companies and researchers indicate that hallucination rates are also rising for reasoning models from companies such as Google and DeepSeek.
Since late 2023, Mr. Awadallah’s company, Vectara, has tracked how often chatbots veer from the truth. The company asks these systems to perform a straightforward task that is readily verified: Summarize specific news articles. Even then, chatbots persistently invent information. Vectara’s original research estimated that in this situation chatbots made up information at least 3% of the time and sometimes as much as 27%.
In the year and a half since, companies such as OpenAI and Google pushed those numbers down into the 1 or 2% range. Others, such as the San Francisco start-up Anthropic, hovered around 4%. But hallucination rates on this test have risen with reasoning systems. DeepSeek’s reasoning system, R1, hallucinated 14.3% of the time. OpenAI’s o3 climbed to 6.8%.
Sarah Schwettmann, co-founder of Transluce, said that o3’s hallucination rate may make it less useful than it otherwise would be. Kian Katanforoosh, a Stanford adjunct professor and CEO of the upskilling startup Workera, told TechCrunch that his team is already testing o3 in their coding workflows, and that they’ve found it to be a step above the competition. However, Katanforoosh says that o3 tends to hallucinate broken website links. The model will supply a link that, when clicked, doesn’t work.
AI companies are now leaning more heavily on a technique that scientists call reinforcement learning. With this process, a system can learn behavior through trial and error. It is working well in certain areas, like math and computer programming. But it is falling short in other areas.
“The way these systems are trained, they will start focusing on one task — and start forgetting about others,” said Laura Perez-Beltrachini, a researcher at the University of Edinburgh who is among a team closely examining the hallucination problem.
Another issue is that reasoning models are designed to spend time “thinking” through complex problems before settling on an answer. As they try to tackle a problem step by step, they run the risk of hallucinating at each step. The errors can compound as they spend more time thinking.
“What the system says it is thinking is not necessarily what it is thinking,” said Aryo Pradipta Gema, an A.I. researcher at the University of Edinburgh and a fellow at Anthropic.
New research highlighted by TechCrunch indicates that user behavior may exacerbate the problem. When users request shorter answers from AI chatbots, hallucination rates actually increase rather than decrease. “The pressure to be concise seems to force these models to cut corners on accuracy,” the TechCrunch article explains, challenging the common assumption that brevity leads to greater precision.
References:
https://www.nytimes.com/2025/05/05/technology/ai-hallucinations-chatgpt-google.html
The Confidence Paradox: Why AI Hallucinations Are Getting Worse, Not Better
https://www.forbes.com/sites/conormurray/2025/05/06/why-ai-hallucinations-are-worse-than-ever/
https://techcrunch.com/2025/04/18/openais-new-reasoning-ai-models-hallucinate-more/
Goldman Sachs: Big 3 China telecom operators are the biggest beneficiaries of China’s AI boom via DeepSeek models; China Mobile’s ‘AI+NETWORK’ strategy
Telecom sessions at Nvidia’s 2025 AI developers GTC: March 17–21 in San Jose, CA
Nvidia AI-RAN survey results; AI inferencing as a reinvention of edge computing?
Does AI change the business case for cloud networking?
Deutsche Telekom and Google Cloud partner on “RAN Guardian” AI agent
Ericsson’s sales rose for the first time in 8 quarters; mobile networks need an AI boost
