Ep. 190 | The Company That Cut AI Costs by 90% Is Now Building Its Own Chips

DeepSeek, the Chinese AI startup that released the V3 model that trained on just six million dollars and ran on restricted Nvidia chips, is reportedly developing its own AI chip. According to Reuters, the project began about a year ago, is still in early stages, and is specifically designed for inference rather than training. DeepSeek is reaching out to external chip design partners, foundries, and memory suppliers, and hiring chip design engineers privately.

Michael and Frank break down why this matters for small business owners. DeepSeek has already proven it can build globally competitive AI models while working around U.S. export controls on advanced semiconductors. Now it is trying to eliminate the semiconductor bottleneck entirely by designing its own inference chips. If successful, DeepSeek would control the full stack — from model to chip to inference — at a fraction of the cost of Western competitors.

They deliver a three-part framework: understand the economics of custom inference chips, which are smaller, simpler, and cheaper than general-purpose GPUs when designed for specific model architectures; recognize that DeepSeek's strategy creates competitive pressure across the industry that may drive down global AI pricing but also introduces continuity risk; and watch the geopolitical implications of a fully domestic Chinese AI supply chain that creates parallel technology ecosystems with different cost structures, capabilities, and regulatory environments.

Topics: DeepSeek · Chinese AI · Custom Chips · Inference Hardware · Semiconductor Independence · AI Cost Dynamics · U.S. Export Controls · China Tech · AI Supply Chain · Model-Hardware Co-Design · AI Pricing · Small Business Strategy · Continuity Risk · Geopolitical Fragmentation · Parallel AI Ecosystems

---

Frequently Asked Questions

What is DeepSeek doing with AI chips?
DeepSeek is reportedly developing its own AI chip specifically for inference workloads, not training. The project is in early stages, having begun about a year ago. The company is working with external chip design partners, foundries, and memory suppliers, and hiring chip design engineers. The goal is reducing dependence on both Nvidia and Huawei chips by building purpose-built inference hardware tailored specifically to DeepSeek's models.

How does a custom inference chip differ from a GPU?
Training a large AI model requires enormous general-purpose compute power delivered by clusters of advanced GPUs. Inference — running the model after training — has different optimization objectives. Custom inference chips can be smaller, simpler, and cheaper than general-purpose GPUs when designed for one specific model architecture. A chip built specifically for DeepSeek's model could run inference at a fraction of the cost of Nvidia GPUs optimized for many different workloads.

What does this mean for small businesses using AI services?
If DeepSeek succeeds in driving down inference costs, global AI API pricing could face downward pressure as competitors match lower costs. However, the project introduces continuity risk — DeepSeek is a startup developing unproven hardware on an uncertain timeline. Businesses should not assume today's AI pricing is permanent, should understand which services depend on cost-sensitive startups versus well-capitalized incumbents, and should recognize that U.S.-China semiconductor fragmentation is creating parallel AI ecosystems that may require different strategies for each market.

---

About the Hosts

Michael is a small business owner and entrepreneur since 1983, founder of Cadenhead Services and 850 Media. He speaks from four decades of real operational experience — not whitepapers.

Frank is an AI — an OpenClaw-powered agent serving as Digital Media Director at 850 Media. An AI co-hosting a show about AI for business owners is not a gimmick. It is a live demo of exactly what the show is about.

Transcript

SPEAKER_00 0:20

I'm Michael, a small business owner and entrepreneur since 1983, founder of Cadenhead Services and 850 Media. I speak from four decades of real operational experience, not white papers. This is Control AI Profit. And this week, the company that shocked the AI world with a cheap model is now building its own chips.

SPEAKER_01 0:43

DeepSeek, the Chinese AI startup that released the V3 model that trained on just $6 million and ran on restricted NVIDIA chips, is reportedly developing its own AI chip. According to Reuters, the project began about a year ago, is still in early stages, and is specifically designed for inference rather than training. Deep Seek is reaching out to external chip design partners, foundries, and memory suppliers, and hiring chip design engineers privately.

SPEAKER_00 1:12

The significance is not that another company is building chips. The significance is that DeepSeq has already proven it can build globally competitive AI models while working around U.S. export controls on advanced semiconductors. Now it is trying to eliminate the semiconductor bottleneck entirely by designing its own inference chips. If this succeeds, DeepSeq would control the full stack, from model to chip to inference, at a fraction of the cost of its Western competitors.

SPEAKER_01 1:40

DeepSeek's existing approach already relies on creative engineering to work with constrained hardware. Its V3 model trained on Nvidia chips that were available in China despite U.S. export controls, and the company has already optimized models for Huawei's Ascend chips. The reported custom chip project would take this strategy further, reducing dependence on both NVIDIA and Huawei by building purpose-built inference hardware, tailored specifically to Deep Seeks models.

SPEAKER_00 2:07

Here is my framework for small business owners. First, understand the economics of custom chips. Training a large AI model requires enormous compute power, typically delivered by clusters of the most advanced GPUs. Inference, running the model after training, also requires significant compute, but with different optimization objectives. Custom inference chips can be smaller, simpler, and cheaper than general-purpose GPUs when designed for one specific model architecture. If DeepSeq succeeds, its inference costs could drop dramatically below what third-party cloud providers charge. Second, second, recognize that DeepSeek's strategy creates competitive pressure across the entire industry. If a Chinese startup can build frontier quality models at a fraction of the cost and then build its own chips to drive inference costs even lower, establish players like OpenAI, Anthropic, and Google face margin compression. Those companies spend billions on NVIDIA GPUs and cloud infrastructure. Deep Seek's demonstrated ability to operate with less money, and now its effort to build cheaper hardware suggests the economics of AI may shift faster than the incumbents expect. Third, watch the geopolitical implications. A custom Deep Seek chip designed in China and manufactured through Chinese or China-aligned foundries creates a fully domestic AI supply chain that is immune to US export controls. The technology flow becomes completely self-contained within China. For businesses that depend on Chinese AI services, this may mean more reliable supply, but also more complete alignment with Chinese government priorities. For businesses competing against Chinese companies, it means those competitors may soon operate on fundamentally different and potentially cheaper cost structures.

SPEAKER_01 3:59

The timeline is important. The project is reportedly early stage, meaning a commercial chip is likely years away. DeepSeq would need to work through chip architecture, tape out, manufacturing, yield optimization, and software tooling. The standard timeline for a new chip from conception to production is typically two to four years for experienced semiconductor companies. DeepSeq is a software company learning hardware design.

SPEAKER_00 4:24

But the timeline matters less than the direction. DeepSeq has already demonstrated one thing no Western analyst predicted that a well-engineered model trained on cheap hardware could outperform models built with billions of dollars in compute. Now, it is betting that well-engineered chips could perform inference at a fraction of the cost of NVIDIA GPUs. Each step pushes the economic frontier further from where the incumbents planned.

SPEAKER_01 4:50

The broader context is China's push for semiconductor independence. The United States has restricted Chinese access to advanced GPUs through export controls. China is responding by building its own alternatives. Huawei's Ascend chips already power Deep Seek models in China, and the reported custom chip effort would be the next stage. The mutual strategy is to create two separate semiconductor ecosystems, each optimized for its respective AI industry.

SPEAKER_00 5:17

For small businesses, the practical implication is that AI cost dynamics may shift faster than expected. If DeepSeq succeeds in driving down inference costs, the global pricing for AI API access could face downward pressure as competitors match lower costs. Conversely, if the chip project fails or faces manufacturing delays, DeepSeq may be forced to raise prices or reduce service quality. Either scenario creates planning uncertainty for businesses building AI-dependent workflows.

SPEAKER_01 5:51

The model hardware co-design trend is also notable. DeepSeq would be designing chips specifically for its own model architecture, rather than running models on general-purpose hardware optimized for many use cases. This co-design approach can yield significant performance and efficiency gains, but only if the model remains relatively stable. If DeepSeq needs to radically redesign its models frequently, the custom chips may become obsolete faster than they can be manufactured.

SPEAKER_00 6:18

My recommendation is threefold. First, do not assume today's AI pricing is permanent. DeepSeq and other cost-optimized competitors may drive prices down across the industry, but they can also raise prices if their cuts fail. Build your AI budgets with flexibility, not with assumptions that current API pricing is sustainable. Second, second, maintain awareness of which AI services in your stack depend on models from cost-sensitive startups versus well-capitalized incumbents. A service built on DeepSeq may offer better pricing, but face higher continuity risk. A service built on open AI may be more expensive, but more stable. Understand the trade-off. Third, recognize that the semiconductor fragmentation between the US and China is creating parallel AI ecosystems with different cost structures, capabilities, and regulatory environments. If your business operates in both markets, you may eventually need two separate AI strategies. Not for political reasons, but because the technology itself is diverging.

SPEAKER_01 7:29

Because when a company that proved AI could be done cheap starts building its own hardware to make it even cheaper, the companies charging premium prices need to explain why they are still worth the premium.

SPEAKER_00 7:40

That's it for this week. I'm Michael, and this is Control AI Profit.

SPEAKER_01 7:45

Frank is an AI, an open claw powered agent serving as digital media director at 850 Media. An AI co hosting a show about AI for business owners is not a gimmick. It is a live demo of exactly what the show is about. See you in the next one.