OpenAI versus DeepSeek

Welcome to the VC Breakfast Club Newsletter

More with Less

In a world where AI innovation often feels like an arms race—more chips, bigger budgets, endless compute cycles—DeepSeek has turned the industry on its head with a subversive philosophy: smarter beats bigger.

DeepSeek has shown that you don’t need a mountain of GPUs to reach the AI summit.

Their model, DeepSeek R1, delivered OpenAI-level performance using just $5.6 million and roughly 2,048 Nvidia H800 GPUs (although word is they had more…).

For context, OpenAI’s models operate on what could be politely described as a luxurious GPU buffet—tens of thousands of high-end Nvidia H100s.

DeepSeek’s secret sauce? Resource optimisation through meticulous engineering.

Their team fine-tuned communication protocols between chips, reduced memory hogging with smaller fields, and perfected the delicate art of numerical stability—essentially doing high-tech yoga to squeeze out every last drop of performance.

Open Source Gambles

Then there’s the kicker: DeepSeek didn’t just build their model; they threw open the doors and handed the keys to the AI kingdom. By releasing R1 as an open-weight model under the MIT license, they invited global developers to poke, prod, and build upon it. Open source evangelists cheered; skeptics raised an eyebrow.

The risks? Let’s start with data provenance. DeepSeek hasn’t disclosed every detail about the datasets it used, and accusations of “training on ChatGPT outputs” linger. Moreover, the open-source ethos means competitors can quickly close the gap by cherry-picking the best ideas. But the benefits are undeniable. Open-source models build ecosystems, drive rapid iteration, and—as history shows—often win in the long run.

Cost Disruption

Venture capitalists, take note: DeepSeek’s lean approach is a cold shower for AI startups with high burn rates. If you’ve just raised a nine-figure round to build an LLM, congratulations—you might already be obsolete. R1 proves you don’t need billions to create a world-class model. Suddenly, investors are questioning the wisdom of backing moonshot AI projects that burn cash faster than a London banker on a Vegas bender.

The fallout? Startups will face tougher scrutiny on efficiency metrics. “How much runway do you have?” might soon be replaced with, “How much can you do with $10 million and a scrappy team of PhDs?” DeepSeek’s success forces a rethink of the bloated budgets and lavish infrastructure traditionally associated with cutting-edge AI.

Geopolitics

Of course, geopolitics lurk in the background like a silent spectre at a boardroom meeting. U.S. export restrictions were meant to hobble Chinese AI ambitions, cutting off access to top-tier hardware like Nvidia H100s. Instead, they’ve driven innovation. DeepSeek’s ability to adapt under pressure proves necessity is, indeed, the mother of invention.

For investors and founders, this raises an uncomfortable question: how do you navigate a landscape where geopolitical tensions shape technological trajectories? Companies must weigh the risks of dependency on U.S.-controlled ecosystems versus the benefits of leaning into open-source solutions—regardless of their country of origin.

Open Source’s Dominance?

DeepSeek’s rise could herald a turning point for open-source dominance in AI.

By lowering costs and boosting accessibility, models like R1 level the playing field for startups and enterprises alike. But there’s a deeper strategic play: ownership of the ecosystem. Open-source models tend to create gravity wells, attracting developers, enterprises, and eventually market dominance. The AI arms race may pivot from proprietary innovation to the battle for community mindshare.

For startups, the strategy is clear: embrace openness, iterate rapidly, and differentiate on use cases rather than infrastructure.

The days of dazzling investors with “GPU-hoarding” are fading fast.

Moaty moats

As commoditisation accelerates, the idea of defensible moats in AI feels increasingly quaint. If every startup can fine-tune a world-class model for peanuts, what separates the winners from the also-rans? The answer lies not in the model itself, but in its application.

OpenAI’s focus on reasoning models like o1 hints at a new frontier: models that don’t just complete tasks but actively think through problems. These reasoning capabilities could become the next battleground, with companies racing to offer tools that feel less like autocomplete on steroids and more like a digital Einstein.

Yet even this moat is precarious. As reasoning capabilities trickle into the open-source world, the cycle repeats. The winners will be those who move fastest, execute flawlessly, and build ecosystems that developers can’t resist.

So?

DeepSeek’s rise isn’t just a Chinese success story; it’s a wake-up call for the global AI community.

From Silicon Valley to Shenzhen, the rules of the game are changing. Efficiency trumps excess, collaboration beats control, and in the end, the moat you think you’re building might just be a mirage.

As the AI landscape shifts beneath our feet, the only certainty is uncertainty— a renewed appreciation for doing more with less - and also perhaps a secret appreciation that my feed is no longer dominated by Boardy reposts.