One week ago today, Nvidia’s market cap crashed by nearly $600B, representing the largest single day drop in history. Most analysts chalked this up to the recent release of DeepSeek’s new reasoning model, which claimed to replicate OpenAI’s newest models efficacy at a fraction of the cost.

The implication? That the entire AI market was a sham, models could be trained for a fraction of what we thought we could, the U.S. had lost it’s lead in the AI race to China, and we were about to witness the most prolific bubble pop since the dot-com crisis.

Obviously, this was a massive overreaction, but nothing generates clicks like fear mongering.

The tsunami of chatter about DeepSeek has everyone questioning the current state of AI, and has inspired every VC on the planet to chime in with their own two cents.

Nobody asked for it, but here’s my take too.

Was DeepSeek’s model release actually a big deal?

Yes and no. Despite the market massively overreacting, there was a fair amount of impressive engineering that went into the DeepSeek model. Being cut off from Nvidia’s best chips meant that DeepSeek’s engineer’s had to innovate in the deepest layers of the software stack, in order to get around access to Nvidia’s much lauded CUDA platform.

This was an engineering marvel that many said could not be done, and is well deserving of praise (CUDA has long been seen as Nvidia’s ultimate moat).

It’s a good example of what a resource starved team can do when pressured to innovate in tight quarters. It’s also a sign that China is not nearly as far behind the U.S. in terms of AI development as we once thought. I would guess that we’re now talking in months, as opposed to years, of a progress gap.

As an actual AI model, though? I’m not sure DeepSeek’s model was all that impressive.

OpenAI has already come out claiming that DeepSeek shameless stole their data from OpenAI (by distilling OpenAI’s latest reasoning model, o1, into their model, r1). This is almost certainly at least partially true (yes, I’m aware of the irony of OpenAI complaining about stealing when they have shamelessly stolen content to train their models for years).

While imitation is the sincerest form of flattery, researchers seem to agree that DeepSeek did not push the frontier of model performance forward. Their model is in no way more performant than OpenAI’s.

If you’re just going to copy someone else’s AI model (or even distill it), instead of pushing the frontier forward, it’s not surprising at all that it could be done for a fraction of the cost.

My takeaways from this whole circus are as follows: i) China is catching up to the U.S. fast, but is still behind; ii) it’s perhaps easier to copy models than we once thought (especially for foreign actors who don’t fear lawsuits); iii) despite the noise, the frontier of model development is unchanged, and iv) the public markets massively overreacted on fear mongering news.

TL;DR: If you thought Nvidia was a buy before this whole fiasco, you should buy the dip. If you thought it was massively overvalued before, it probably still is.

What does all of this mean for AI startups?

Despite the chicken little event, this much attention has caused many large enterprises and investors to reassess their view of the broader AI market.

If models are easier to copy but harder to improve upon than ever, what does that mean for AI startups in 2025?

A few of the perspectives we’re consistently hearing from well informed researchers and founders are as follows:

  1. Only the biggest labs are going to consistently push the AGI frontier forward: Despite what anyone says, at this point I think pushing the model frontier forward is big tech’s game to lose. It requires too many resources to justify chasing without a massive distribution base upon which you can capitalize on your investment (at this point, I think the real players are Google, OpenAI / Microsoft, and Anthropic / Amazon. That’s it).

  2. Selling models alone isn’t a good business: If foreign actors are going to steal and then open source models, prices are going to be driven down to zero. Many have questioned this strategy for ages, but it’s now being proven out in real time. Gone is the dream of building a great business by just selling an API via token-based pricing (which, as all analysts have noted, was never profitable anyways).

  3. All the AI labs are increasingly becoming full-stack product companies: Instead of selling API access to their models, all of the best AI companies are increasingly moving closer to the customer and their specific use cases. Last mile delivery for these companies is becoming increasingly important as the killer apps drive the businesses (e.g., ChatGPT and Claude have become the driving forces of growth at OpenAI and Anthropic, respectively).

  4. If you’re not focused on pushing the AGI frontier forward, you can replicate best-in-class performance for a fraction of the cost: Similar to what the DeepSeek team did, you can train best-in-class models for significantly less money if you only focus on doing what you know works (even without blatantly stealing from OpenAI). This is particularly true in the Enterprise world, where models can leverage RAG to be much smaller while still being just as performant (we’ve already seen this play out at both Cohere and Writer).

  5. As best-in-class models get open-sourced, there’s never been a better time to build AI-native Applications: As model costs collapse, a tsunami of startups are getting access to near world-class performing models for almost nothing. This is already leading to a golden era of productivity gains and business creation, which we don’t expect to slow down anytime soon.

The last point is by far the most important. A few months ago we spoke about entering into the era of Actual AI deployments. This has only continued to accelerate, as more and more enterprise buyers are eager to adopt new solutions. Not all solutions are being adopted equally, though. What does this mean?

My prediction is that 2025 will be the year of non-bullshit AI.

Buyers are sick of confusing and vague AI promises, and are looking for solutions that just work. In other words, solutions that are specific, understand industry-specific pain points, and clearly communicate value.

I (like many others) am tired of the “wrapper or not-a-wrapper” debate. Companies that deeply understand their customer needs are being rewarded. Companies that expect their customers to come to them are not. That’s it.

While having a strong AI team that can keep a finger on the pulse of the latest innovations is still going to be very important, the reality is that AI Applications are really product-driven companies.

If nothing else, I expect 2025 to be a massive year for shipping incredible products. Models are more than good enough to create transformative product experiences already.

This is the year that I expect knowledge workers to begin adopting task specific productivity tools in droves (whether they be copilots, agents, swarms, or whatever the buzzword of the moment is).

All of us should see the quality of our jobs improve over the next twelve months.

It’s up to the next generation of Entrepreneurs to apply these new AI capabilities to problems in novel, creative ways. The reward of unbound riches waits at the other end, but competition will undoubtably be fierce.

The race is on.

Subscribe now

Leave a Reply

Sign Up for TheVCDaily

The best news in VC, delivered every day!