DeepSeek V4: Low-Cost Scale and China’s Answer to US Foundation Models
DeepSeek’s new V4 Flash and V4 Pro models underline how a Chinese AI startup is competing head-on with US giants while betting on low-cost scale. The DeepSeek V4 model family is open-weight and text-only, with both versions offering context windows above 1 million tokens, allowing massive documents or full codebases to be processed in a single prompt. V4 Pro uses a mixture-of-experts architecture with 1.6 trillion parameters and 49 billion active at once, which DeepSeek says makes it the largest open-weight model currently available. It significantly exceeds the parameter counts of rival Chinese systems from Moonshot AI and MiniMax, and more than doubles DeepSeek’s own V3.2 model. The company claims that the top-end “V4 Pro Max” achieves superior reasoning performance versus OpenAI’s GPT‑5.2 and Google’s Gemini 3.0‑Pro, while only slightly trailing their newer flagships. V4 also leans on Huawei chips, signalling a deliberate pivot away from Nvidia and other US suppliers.
From R1 to V4: Performance, Cost and a Brewing Data-Scraping Backlash
DeepSeek first shook markets with its R1 reasoning model, which showed that an open-weight system could rival frontier closed models while running on less advanced Nvidia H20 GPUs. By combining distillation, mixture-of-experts and multi-head latent attention, the company extracted more performance per unit of compute and positioned itself as a cost-efficient challenger in the China AI industry. That playbook continues with the DeepSeek V4 model lineup: open-weight, long-context and explicitly marketed as more cost-effective than Western peers. Yet its rise has been shadowed by accusations from Anthropic and OpenAI that DeepSeek unfairly built its systems using their models and data, amplifying wider concerns about AI data scraping and “AI theft.” The long-anticipated V4 launch, which some analysts expected months earlier, is therefore being read in two ways: as evidence of China’s technical progress and as a flashpoint in growing disputes over how training data is gathered, shared and monetised across borders.
DeepRoute Self Driving: Mass ADAS Deployment as a Strategic Beachhead
If DeepSeek represents China’s foundation model ambitions, DeepRoute.ai embodies the country’s push to industrialise AI in physical systems. The autonomous driving technology developer says more than 300,000 vehicles on Chinese roads now use its advanced assisted driving system, with CEO Maxwell Zhou projecting that figure could increase by another 1 million vehicles this year. Rather than focusing solely on fully driverless robo-taxis, DeepRoute’s strategy leans on large-scale deployment of practical ADAS capabilities through automaker partnerships. This positions DeepRoute self driving as a mainstream feature embedded in mass-market cars, not just premium flagships or small pilot fleets. The company’s momentum highlights how rapidly AI-enabled driving functions can spread once automakers standardise on a software stack. It also underscores a core element of China’s AI strategy: use huge domestic markets in autos and mobility to generate real-world data, refine models and lock in ecosystems before regulatory or geopolitical headwinds slow overseas expansion.
China’s AI Strategy: Scale, Cost and the Shadow of ‘AI Theft’ Allegations
Taken together, DeepSeek and DeepRoute show how the China AI industry is pursuing a dual track: aggressively improving foundation models while pushing applied systems into hundreds of thousands of vehicles and other devices. The emphasis is on competitive performance at lower cost, open-weight releases that attract global developers, and large-scale deployments that generate feedback loops of data and usage. Yet the same dynamics raise thorny questions about AI data scraping, model provenance and cross-border regulation. Allegations that DeepSeek relied improperly on rival models sharpen Western concerns over IP protection and the difficulty of auditing what goes into large models. For Chinese AI firms, this could translate into tighter export controls, restrictions on cloud access and hesitancy from foreign partners who fear regulatory or reputational risk. The resulting uncertainty may push Chinese companies to double down on domestic and Global South markets where scrutiny is less intense but competition is also increasing.
What Global Users Should Watch in the US–China AI Rivalry
DeepSeek’s V4 launch within hours of OpenAI’s latest release, and DeepRoute’s rapidly expanding ADAS footprint, illustrate how the US–China AI rivalry now plays out in near-synchronised product cycles and deployment races. For global users and partners, the key questions will be less about headline benchmark scores and more about trust, governance and supply chain resilience. Enterprises considering Chinese AI startup offerings must weigh the appeal of open-weight, low-cost models against potential exposure to sanctions, IP disputes or data localisation rules. Policymakers will scrutinise whether alleged “AI theft” is a one-off controversy or a systemic issue that justifies new global rules on training data and model transparency. Meanwhile, automakers and platform companies must plan for a world where AI stacks are increasingly geopoliticised. Watching how DeepSeek V4, DeepRoute self driving and their Western rivals evolve will offer early signals of whether the AI ecosystem fragments or finds a workable middle ground.