How Chinese Companies Are Taking Over The U.S.

China's largest coffee chain, is taking on the U.S.

At the beginning of September 2025, just two months since its launch, Luckin had opened five stores in New York City.

Based on its track record in China, Luckin could absolutely pose a threat to Starbucks dominance in the U.S.

Welcome to the world of micro dramas.

Over the top storylines.

Could he be someone dangerous?

Jaw dropping twists.

He helped save my company.

All packed into snack-sized, vertical video clips.

In the States, you have soap opera, you know, 'The Bold and the Beautiful,' and it's 'The Bold and the Beautiful' on steroids, basically.

You've got to have a cliffhanger on every minute versus every hour.

So, everything is sweet, short and sharp.

Let's talk about DeepSeek because it is mind blowing and it is shaking this entire industry to its core.

Microsoft and OpenAI are investigating whether their new Chinese competitor used the OpenAI model to train its rival chatbot.

China's largest coffee chain is taking on the U.S.

At the beginning of September 2025, just two months since its launch, Luckin had opened five stores in New York City.

The coffee giant has only been around since 2017, but it's become a dominant force in its domestic market with almost all of its 26,000 locations in China.

For reference, Starbucks has just under 8,000 in China and about 17,000 in the U.S.

CNBC visited one of the first luck and locations in New York. It had a number on the coffee counter,

New York. It had a number on the coffee counter, 00002. The zeros suggest that Luckin may be looking

00002. The zeros suggest that Luckin may be looking to scale its locations here into the thousands.

But, what challenges could Luckin face as it enters the U.S. market, and what kind of threat could it pose to

U.S. market, and what kind of threat could it pose to coffee chains like Starbucks?

Luckin is a relatively young company, but has already experienced some major ups and downs.

Founded in June 2017, within two years the company went public on the Nasdaq.

Its initial growth was rapid and Wall Street was bullish on its future. From Q1 of 2018 to Q1 of 2019, Luckin's customer base soared from 485,000 people to 16.9 million.

That's about 45,000 new customers per day.

It's a volume game.

At the end of the day, Luckin is pointing to have the sales leverage come through the door, minimize the cost of the real estate by having as many transactions as possible, and so be it if the marginal profit on the marginal coffee is relatively small.

They are trying to optimize for transactions.

But, only after 18 months as a publicly traded company, Luckin was charged with fraud by the SEC in its preliminary IPO filing.

The company noted that it is an emerging growth company, which meant it was eligible for reduced reporting requirements. Most notably, Luckin didn't need an

requirements. Most notably, Luckin didn't need an auditor to attest to its internal management.

In January 2020, short seller Muddy Waters Research shared an anonymous report that alleged fraudulent reporting of key business metrics to manipulate investor confidence.

The report said it utilized 92 full time and over 1,400 part-time staff to run surveillance and record store traffic for 981 days in 620 stores.

Luckin's stock price tanked in the wake of the report.

In February 2020, the company issued a press release calling the short sellers report misleading and false. Investors and analysts also had doubts

and false. Investors and analysts also had doubts about its validity, and the stock rebounded.

But, by April, an internal investigation found the COO had fabricated 2019 sales by $310 million.

Shares fell over 80%, wiping out $5 billion off its market cap. It turns out the company was also reporting millions in profits when restated filings show it was actually operating at a loss.

Luckin was delisted from the Nasdaq in June 2020 and filed for bankruptcy less than a year later.

But, Luckin emerged from bankruptcy in 2022.

It brought in new leadership and cleaned up its balance sheet over the next three years.

It more than tripled its store count and quickly overtook Starbucks in China by total revenue.

Despite a report from the Financial Times that Luckin was plotting its return to the Nasdaq, the company still trades on the OTC market, or over-the-counter market, a less regulated exchange.

Its share price is up around 100% over the past year.

Luckin relies entirely on mobile ordering.

It cuts down the wait time in the stores and also reduces the amount of labor that it needs to operate.

I see a coupon here for $1.99 drink on my first order. Luckin's full priced items aren't actually that

order. Luckin's full priced items aren't actually that far off from Starbucks, but the difference is you're rarely ever paying full price.

The app showers you with coupons.

So much so that it's actually unlikely its initial stores are even profitable right now.

That is not how its competitor operates.

Starbucks has always strived to be profitable on a single occasion. In the case of Luckin,

occasion. In the case of Luckin, the idea is 'I want to grow in awareness.

I want to make sure that the brand gets recognized on a national basis, even though at the beginning this means that I might need to be suffering from some smaller losses on a per store basis.'

New York is an expensive place to operate a coffee shop. Equity research firm Bernstein analyzed its

shop. Equity research firm Bernstein analyzed its initial profitability.

Rent for Luckin's Midtown location is around $15,000 per month and labor costs are around $66,000 per month. Utilities, maintenance and insurance

month. Utilities, maintenance and insurance could cost another $10,000.

That puts its overhead at around $92,000.

In the report published in late July 2025, less than a month after opening, the stores were estimated to be generating $85,000 per month, averaging about 5 to 600 orders per day.

Order volumes would need to be double that to break even.

I ordered on the app to test out the experience.

It says it should be ready in just a couple of minutes.

Walking up to the store.

Here is Luckin.

Thank you.

Thank you.

Luckin had some really interesting and innovative drinks in there. I saw pineapple cold brew on the menu. This is a coconut iced latte.

menu. This is a coconut iced latte.

It's one of their most popular drinks and it's really good. It's just a hint of coconut and not too

really good. It's just a hint of coconut and not too sweet.

The company rolled out nearly 120 new drinks and food items in China in 2024, but it launches these new varieties carefully, using a strategic, data-driven approach to test new products on the market.

Neither Starbucks nor Luckin agreed to an interview with CNBC for this story, but the chinese coffee company had noted in a press release that its initial soft launch in New York City would enable the company to gain localized operational insights into site selection, product innovation and customer experience, providing valuable insights for the company's global expansion.

Traffic to the Starbucks across the street from the Midtown location that we visited was not particularly affected during the initial weeks of its launch.

This specific Starbucks was already seeing large year-over-year drops in visits, irrespective of the Luckin opening.

Replicating Luckin's success in China here in the U.S.

may not be as easy as it seems. Here in the United States, the coffee culture is much more mature and a new brand, for as a new challenger it might seem, they are facing some local competition to national competition, and so displacing some of the larger players might be just harder.

Secondly, purely from a cultural standpoint, especially in this day and age, we are seeing the United States retrenching a little bit more toward American brands.

And, while many chains are rewarding customers for using their mobile app.

Some might be adverse to that as their only option.

The woman in front of me tried to add an additional drink to her order in person, and the guy said, 'No. You know, it all has to go through the app.' The guy behind me walked in, and I guess he didn't realize that it was an entirely mobile, cashier-less environment, and he tried scanning the QR code, downloading the app, got frustrated and just walked out. So, there still seems to be a

walked out. So, there still seems to be a little bit of friction around this idea of an entirely mobile environment.

It's distinct from Starbucks, which is telling you sort of coming into stores that employee connection is still important.

Luckin doesn't necessarily have that, right? So, I

think our view would be that that isn't a broader threat to Starbucks, do they establish somewhat of a foothold in major cities.

But, the past couple years have been rocky for Luckin's main competitor, Starbucks.

The company brought in former Chipotle CEO Brian Niccol last year to revitalize the brand.

The higher cost the company $85 million in cash and stock compensation.

He's made all sorts of top-down efforts, like having baristas write messages on the cups and adding more comfortable seating into the stores to revive the brand's coffeehouse image.

In the case of the United States, for Luckin will be trying to approach the oversea expansion without necessarily having the same level of, uh, localized and deep-rooted experience into this market that enable them to be scaling rapidly in China.

I'm not saying that this may not be a threat to Starbucks, but Starbucks has still the incumbent.

They still are significantly more relevant for a U.S.

consumer. But, what we have seen in the store might be indicative of the fact that if the awareness were to be rising, consumers might be lacking that product and could be a potential threat to Starbucks.

Do you need to cut price at all? You talk about the value scores given there are so many new entrants into the beverage space with so much added competition.

McDonald's, Luckin, there are so many others.

Are you thinking about price cuts?

I do believe we are a premium brand, and I do believe you get a premium experience with the green apron service model, and what we're doing to our coffeehouses to give people that coffeehouse experience again. So, I think we're priced correctly right now,

again. So, I think we're priced correctly right now, and I like our competitive position.

Where Luckin shines is value.

Its products are not actually that far off from Starbucks prices.

A latte at the Luckin we visited was $5.75.

At Starbucks, it's $5.95.

The key difference, however, is that customers are rarely paying that full price at Luckin.

They are showered with coupons in the app, often in the range of 30% to 50% off.

In 2024, Starbucks began offering frequent deals and promotions within its app, too.

This attracted customers that were more price sensitive, but Nichol had the company move away from that.

It was driving away the core consumer base, who was feeling fairly disappointed by the lower quality of experience despite the premium price that they would need to pay for Starbucks coffee.

The Starbucks is a premium brand, and by virtue of being a premium brand, you attract consumers that, on average, tend to be a little bit more price insensitive, a little bit less attracted by discount.

And so, they can protect their margins a little bit better than a company that is mostly known for their discount models.

Based on its track record in China, Luckin could absolutely pose a threat to Starbucks dominance in the U.S., but this is Starbucks' home turf, and despite some recent speed bumps, the company has managed to scale its operations here profitably for years.

Plus, the coffee market is highly saturated in the U.S.

with thousands of local players.

Now, only time will tell just how big of a disruptor Luckin will be.

I only owe my success to that secret CEO of King's Corp..

Corp..

You will be groveling at my feet for forgiveness when you realize who I really am.

Spoiler alert. That man earning a meager wage does, in fact, turn out to be the secret CEO of King's Corp.

This is one of the micro dramas featured on ReelShort, a Chinese-backed, short video app.

I'm willing to wager my entire reputation that this man is a useless loser.

Welcome to the world of micro dramas.

Over the top storylines.

Could he be someone dangerous?

Jaw dropping twists.

He helped save my company.

All packed into snack-sized, vertical video clips.

Fast, addictive and ready to take over your feed.

In the State to have soap opera, you know, 'The Bold and the Beautiful,' and it's 'The Bold and the Beautiful,'on steroids, basically. You've got to have a cliffhanger on every

basically. You've got to have a cliffhanger on every minute versus every hour.

So, everything is sweet, short and sharp.

They're tapping into the behavior of instant gratification.

The way that you immediately are satisfied by seeing this story is what is the hook.

It's a little over the top, but it's over the top that it's still entertaining, right?

China's micro drama market exceeded ¥50 billion, or approximately $6.9 billion USD in 2024, outpacing China's own box office.

And, it now has its sights set on the U.S.

market.

Many micro drama producers believe, because America produced Hollywood and is probably the world's largest cultural export, if you can sell it to and make it appealing to the American public, you can then sell it and make it appealing to the rest of the world.

So, what is behind the popularity of China's micro dramas? And can the nearly $7 billion industry become

dramas? And can the nearly $7 billion industry become China's next big cultural export to disrupt the U.S.

entertainment industry?

Micro dramas started in China, I believe, in around 2018, and they were sort of an offshoot and a way to capitalize on the trend and the popularity of short-form video, which was made popular by TikTok.

The format itself has become increasingly popular amongst users. It's where their time is spent,

users. It's where their time is spent, which is why you saw Meta introduce Reels to Facebook and Instagram, and why YouTube introduced Shorts, to capitalize on this short-form video content.

In China, even though they have iQiyi and Tencent, they don't have Netflix. They don't have a lot of the Western apps, they don't have, so they have to create their own.

It all started with Douyin, and also Kuaishou, as they started pushing this vertical, short dramas on their platforms. Especially 2020, 2021, during the pandemic, dramas grew rapidly.

Low-cost production, a fast-paced nature, about 90s to 120s long, so whatever you have fragmented time you watch.

It is not really about the quality, the cast, the story.

It's really about how effective you are to grab the audience within that second.

And, that's what gets people engaged because it almost goes right into the crux of the story, wherever the drama is.

It doesn't it require any effort on the part of the user to just jump in and start to watch.

These are like your telenovela, basically. When you look at the genres,

basically. When you look at the genres, a lot of romantic comedy.

They have the vampire stories.

They have, like, the tycoon stories.

Unlike traditional big budget TV series, dramas offer producers a low-cost, high-reward format that's fast to make and easy to scale.

The streaming companies, a huge cost center for them is content creation.

For a short-form drama, you can get intriguing content at a much lower cost with a shorter production time, so you can have more volume.

The leading apps can turn 8 to 10 shows a month.

You're looking at, like, producing 100 original micro dramas per year.

They are usually taking unknown actors to play these parts.

You're looking at anything from top to finish within two months versus a year and a half.

China's micro drama market surpassed the film industry's annual box office revenue for the first time ever in 2024. Part of micro drama's growth is also attributed to the marketing strategies geared towards their intended audience, typically women between the ages of 25 to 35.

They are able to produce it precisely, but also actually direct the marketing to the audience.

Everything is on data.

It's not just about content, it's the precision in production and marketing, which is, I think, is genius, actually. Major players actually wanted to go beyond

actually. Major players actually wanted to go beyond China. The first place they looked at is U.S.

China. The first place they looked at is U.S.

Downloads of micro drama apps in the U.S.

reach 10 million in April of 2025, an increase of 150% year-over-year, while monthly active users surged over 300% over the same period.

ReelShort, DramaBox and GoodShort were the top downloaded apps in the U.S.

for micro dramas.

They accounted for about 50% of the downloads in year-to-date 2025.

ReelShort, DramaBox and GoodShort were unavailable to accommodate CNBC's request to participate in the story. The U.S.

the story. The U.S.

is the world's largest revenue generating market for micro drama apps, contributing 60% of global mobile revenue in 2024.

In addition, the three short drama apps rank among the top 20 media and entertainment apps in the U.S. by mobile revenue so far this year.

U.S. by mobile revenue so far this year.

So, the U.S. is very important to micro drama apps because it is a higher ARPU, average revenue per user, market.

So, that means that they can charge more per user.

The cost of the app, when they monetize it, is higher in the U.S.

With the popularity of micro dramas steadily rising in the U.S., the question becomes whether the nascent

the U.S., the question becomes whether the nascent format can make a dent in the U.S.

entertainment industry and pose a challenge to existing players in the space.

In my view, micro dramas do not pose a direct threat to a platform such as Netflix or traditional TV.

At least, not yet.

They kind of serve different purposes.

You watch a program on a streamer for the plot, the set, the acting, whereas this, I think, is just like a quick bit of entertainment.

But, I wouldn't say that, 'Oh, is it gonna take away the movies? Is it gonna take away Netflix?' I wouldn't

the movies? Is it gonna take away Netflix?' I wouldn't say that, but I think definitely something to watch out for.

Micro dramas are not aiming to become those big budget shows or films. Instead, they are looking to carve out that unique space by targeting, you know, casual, short attention span viewers. Micro dramas will pose more threat to social

viewers. Micro dramas will pose more threat to social media companies because at this stage they are competing for the eyeballs.

And, I can either choose watching dog videos on TikTok or I can watch micro drama.

Typically, not both.

Some have compared the rise of micro dramas to that of Quibi.

Short for quick bites, featuring content under ten minutes designed for viewing only on your phone.

A similar short-form mobile streaming service that shut down in just six months after it launched in 2020.

During Covid, everybody was at home.

It had a lot more time and it was just really, really bad timing for them.

Hollywood-style content is something that they try to replicate, but in a very short, kind of, fast-paced environment.

You know, vertical is popular.

Attention is becoming shorter.

They saw it. But, when it comes to execution, a lot of things went wrong for them.

Experts say it's more likely these short-form dramas will coexist alongside other media and entertainment companies.

Would I be worried as a social media platform?

Yes, because somebody is taking away my eyeballs.

Would I be worried as Hollywood, Netflix and other platforms?

Yes, but not because they are the threat.

Its because the behavior, and the market that they had created, is a threat.

You cannot compare street food to Michelin.

Is Michelin food is going to take over street food? No,

not a chance. But, there is an appetite for street food. There is an appetite for Michelin

street food. There is an appetite for Michelin dishes. So, that's all I can sort of compare between

dishes. So, that's all I can sort of compare between micro drama and long-form drama.

The rise of dramas could force existing media companies to rethink their content strategies going forward.

Given the strength of the short-form video format, I think that this will continue to grow, similar to the way short-form video took off, and it forced Meta and Google to add Shorts and Reels.

What I think will be very interesting is for the streamers to really pay more attention to the vertical audience, and then perhaps to produce more to cater to them, so that when they are actually don't have time to watch 5 to 6 hours, then they can still engage the audience.

They also become more attractive partners to these other players to do partnership deals.

A lot of great IPs from the studios can be repurposed.

If we adapt into micro drama, it becomes like a pretty interesting ecosystem.

There are examples in China where collaboration with social commerce, or e-commerce, are very successful, or collaboration with a more traditional, long-term video creator are successful.

It can be a very exciting medium to play together with the long-form. People keep saying,

the long-form. People keep saying, you know, TV is dying.

It's not dying. It's just moving to the verticals, moving to the fragmented viewing.

So, if people moves, so you have to move with them. Otherwise, you'll be eliminated.

them. Otherwise, you'll be eliminated.

$1 trillion sell off.

Front and center this hour. The deep sell off.

Major tech sell off today.

DeepSeek off is still really pressuring tech.

Triggered on its face by one Chinese startup.

Let's talk about DeepSeek because it is mind blowing and it is shaking this entire industry to its core.

But underneath, stemming from growing fears around a technique that could upend the AI leaderboard.

A technique called distillation.

Distillation.

Their work is being distilled.

Distillation is the idea that a small team with virtually no resources can make an advanced model by essentially extracting knowledge from a larger one.

DeepSeek didn't invent distillation, but it woke up the AI world to its disruptive potential.

I'm Deirdre Bosa with the TechCheck take.

AI's distillation problem.

AI models are more accessible than ever.

That is the upshot from a technique in AI development that has broken into the main narrative called distillation. Geoffrey Hinton,

distillation. Geoffrey Hinton, dubbed the godfather of AI, he was the first to coin the term in a 2015 paper while working as an engineer at Google. Writing that distillation was a way to

Google. Writing that distillation was a way to transfer the knowledge from a large, cumbersome model to a small model that's more suitable for deployment. Fast forward to today,

for deployment. Fast forward to today, and upstarts, they're using this method to challenge industry giants with years of experience and billions in funding.

Put simply, here's how it works.

A leading tech company invests years and millions of dollars developing a top-tier AI model from scratch. They feed it massive amounts of data,

scratch. They feed it massive amounts of data, harness huge amounts of compute and fine tune it into one of the most advanced models on the market. Then, a smaller team swoops in.

market. Then, a smaller team swoops in.

Instead of starting from scratch, they train a smaller model by using a technique called knowledge distillation.

Pummeling a larger model with questions and using that to train its own smaller, more specialized model.

By capturing the advanced model's reasoning and responses, they create a model that's nearly as capable, but much faster, more efficient and far less resource intensive.

This distillation technique is just so extremely powerful and so extremely cheap, and it's just available to anyone, anyone can basically do it, that we're going to see so much innovation in this space on the LML layer.

And, we're going to see so much competition for the LLMs, so that's what's going to happen in this new era that we're entering.

Tech giants use the technique, too.

In fact, Google was a pioneer in distillation, thanks to Hinton's research.

Just weeks before DeepSeek broke onto the scene, Google was already using distillation to optimize lightweight versions of its Gemini models.

Once again, Google had the tech, but someone else, DeepSeek this time, turned it into the story.

Just like Google pioneered transformers that made generative AI possible, only for OpenAI to swoop in with ChatGPT and own the narrative.

DeepSeek showed Wall Street just how effective distillation could be.

People talk a lot about DeepSeek, and these new models that seem to be doing the work that was done in years before, or months before, in days. People need to remember what's happening is

in days. People need to remember what's happening is they're distilling, which basically means building, on the frontier models that have been created.

Able to mimic, and even surpass OpenAI's advancements, in just two months spending what it says was less than $6 million on the final training phase.

But, it's too simplistic to attribute its success to just copying.

It became increasingly clear that they had made some fundamental improvements in the way to approach this.

DeepSeek also applied clever innovations.

If distillation was the only reason China got these models, like DeepSeek got these models, then Microsoft should have also gotten them, right? Like there's

something more to it.

And, it's not just compute.

It's not just distillation. It's not just access to, like, a lot of tokens.

It's about cleverness.

Yet, distillation was a major factor in DeepSeek's rapid ascent, helping it scale more efficiently and paving the way for other, less capitalized startups and research labs to compete at the cutting edge faster than ever before.

There were these researchers at Berkeley, just a week ago, showed that for $450 bucks, they could create models that were almost as smart as this reasoning model from OpenAI called O1.

For $450, that's not $450,000.

So, just $450 bucks.

They could do that. So, this distillation really works.

That new reasoning model, called Sky T1, made by distilling leading models in just 19 hours with $450 and just eight Nvidia H100 chips.

There's a little small detail in the technical report in the research paper by the folks.

They actually took an existing model. It's called

Qwen. This is not a reasoning model.

It's an older model by Alibaba.

And, they said, 'hey, can we make that smarter so that it can do reason?' One approach they had was to do this kind of reinforcement learning, this kind of fancy technique. They did that, and it was much better. Qwen got much better, but they also just tried distilling it.

So, they just took 800,000 outputs from R1, and then they just did very basic tuning, very cheap, simple tuning, and that actually beat all of the other approaches.

Just weeks later, researchers at Stanford and the University of Washington created their S1 reasoning model in just 26 minutes, using less than $50 in compute credits. And Hugging Face,

compute credits. And Hugging Face, a startup that hosts a platform for open source AI models, recreating OpenAI's newest and flashiest feature, deep research, just for fun.

Hosting an in-house, 24-hour challenge that resulted in an open source AI research agent called Open Deep Research.

Which begs the question, why are big tech firms still investing billions of dollars to push the frontier, developing the most advanced AI models when someone can just turn around and distill it for significantly less work and less money?

That Microsoft and OpenAI are investigating whether their new Chinese competitor used the OpenAI model to train its rival chatbot, including unauthorized access to OpenAI developer tools.

Also, David Sacks, Trump's AI and crypto czar, said on Fox yesterday, this quote substantial evidence DeepSeek quote distilled knowledge from OpenAI models to create its own products.

If model distillation is accelerating, which these recent developments show is happening, and any small team can leap ahead of the biggest AI companies, what's happening to their competitive edge?

Another key paradigm shift post-DeepSeek.

The rise of a new open source order.

A belief that transparency and accessibility drive innovation faster than closed door research.

My joke is everybody's model is open source, they just don't know it yet, you know, because it's so easy to distill them.

So, you might think you haven't open sourced your model, but you actually have.

DeepSeek open sourced its V3 and R1 models, publishing essentially a blueprint that allows anyone to download, use and modify it for free.

Likewise, those models published by university researchers and Hugging Face they were all open source too, raising questions about OpenAI's closed-source strategy that once seemed like a moat, but suddenly looked more like a target.

Open source always wins.

You know, like in tech industry, you know, that has been like one truth, you know, of the last three decades is that you cannot beat the momentum.

Even OpenAI CEO Sam Altman himself walked back his closed-source strategy.

In response to a question on Reddit whether OpenAI would release model weights and publish research, Altman replied, 'Personally, I think we have been on the wrong side of history here and need to figure out a different open-source strategy.' Just a remarkable statement from a leader,

strategy.' Just a remarkable statement from a leader, who has long championed the closed-source approach, citing safety, competitive dynamics and monetization as key justifications.

For Altman to now admit that OpenAI may have been on the wrong side of history, that suggests mounting pressure and could have ripple effects across the AI landscape. Distilled models,

landscape. Distilled models, they can be created on the cheap, and without evaluation to protect, the universities and startups creating the models, they seem inclined to give it away for free.

The biggest winners?

Developers.

The open source keeps the proprietary players in check from a pricing and performance standpoint.

Because if I know, as a developer, that I can always run an open-source model on my own infrastructure, then that really reduces the pricing power of those proprietary players, which again, is a huge win.

The cost of running AI applications is plunging.

DeepSeek's R1, for example, cost $2.19 per million tokens.

OpenAI's comparable O1 costs $60.

So, someone building an AI app would have to pay multiples more to use OpenAI instead, which means the script has now flipped.

Teams building AI applications, they used to be disparagingly called ChatGPT wrappers, accused of having no competitive edge since their entire AI app, or website, is just an interface on top of an OpenAI or Google or Anthropic model.

But, with the cost to serve steadily declining, those AI application makers, they now have an advantage.

Every time you make AI more efficient, you actually open up a dramatic increase in more use cases.

I think if you zoom out and you say, if AI became 5x more efficient, or 10x more efficient, I'd argue that we'll have 100x more use cases in the next 5 to 10 years.

So, I think this is a win for developers.

I think it's a win for anybody building at the application layer of AI, and I think it's a win for the long-term AI ecosystem to have a continued sort of deflationary effect of the cost of these models.

And, AI model builders who have been on top for the past few years, they're looking more and more commoditized.

What distillation has not done is change the calculus for the biggest players in AI.

OpenAI is on track to raise $40 billion from SoftBank.

Meta Microsoft Google Amazon they all said, on their first earnings report post-DeepSeek, that not only were their AI spending plans intact, but they would ramp up capital expenditures, or capex, even more. Nvidia recovered from the DeepSeek sell off,

even more. Nvidia recovered from the DeepSeek sell off, and Jevons Paradox became the leading narrative.

There's plenty of people who are quoting Jevons Paradox this morning, talking about the fact that an increase in efficiency for a particular resource leads to more use of that resource.

Are you in that camp as well?

Absolutely. I believe fully that these new, innovative techniques will lead to the development of more models, more testing and will lead us towards AGI faster.

Smaller, more targeted models are optimized for speed and efficiency, making them cheaper to use, not just for developers, but also for enterprises integrating AI into their businesses.

Here's Tuhin Srivastava, founder of AI infrastructure company Baseten.

What we are seeing from our customers is that when it comes to using these models in production, oftentimes, those smaller models are good enough, and now you're taking that really big power and using it in that smaller model.

He says he heard from nearly two dozen Fortune 100 companies in the week following DeepSeek's reasoning model breakthrough.

Not only do they want to run more efficient models, but it's changed their view on licensing AI.

They're questioning the premium they pay for OpenAI's ChatGPT APIs in Azure and Anthropic's Claude in AWS. Which brings us to OpenAI.

in AWS. Which brings us to OpenAI.

The pressure is on.

And for Altman and his team, the holy grail is AGI, or artificial general intelligence. For them, and the best capitalized

intelligence. For them, and the best capitalized players, you could call it an AGI at all-costs strategy. The idea that they will continue to push the

strategy. The idea that they will continue to push the technological frontier relentlessly, and reaching it first is worth any investment, rather than relying on incremental improvements.

Half $1 trillion for Project Stargate, all part of Sam Altman and SoftBank CEO Masayoshi Son's grand ambition to reach artificial superintelligence.

AI that surpasses human intelligence in all aspects, including creativity and problem solving.

Distillation may yield more cost effective performance gains, but it does not drive the revolutionary breakthroughs needed to reach AGI or ASI.

And, that is why the race has never been more urgent.

In a world where AI capabilities can be distilled, refined and replicated faster than ever, the window to build a true frontier level advantage is narrowing by the day.

These topics we cover in the TechCheck Take.

They're complicated, and we heard your feedback on our long-form interview within our DeepSeek piece.

So, this time we're bringing you an in-depth conversation with another pioneer in the space, Glean CEO Arvind Jain.

He's a Google veteran, and he's not just observing distillation's impact, he's applying it in real-world, enterprise AI solutions.

Thanks for watching and we hope you enjoy.

Arvind, thank you so much for being here. Let's talk

about some of the changes that have just happened in this broader AI landscape.

I saw you maybe a month, or two months ago, we got DeepSeek.

We got sort of questions around scaling laws.

What do you think the biggest takeaway from what's what DeepSeek's breakthrough was?

Well, I think the big thing is between 2024 and 2025, there's a big shift in how the industry thinks what's going to happen next with AI models.

Like last year, we were all thinking that model building is reserved for these, the largest companies out there, and there's going to be only 3 or 4 of them.

And, now you're seeing that, no, that's not actually the case. Like, you know, you can have, like, you know, companies that can come and build amazing models, like, you know, which are comparable to the state of the art with much lower training costs, and there's going to be so many of them now.

And also, the industry has sort of also like these models are no longer like... There's no longer like the best model out there.

That concept has gone away. Like we had that like two years back, of course, but now different models are all getting better at different things. And for

you as like, you know, for your use case, like if you're an enterprise trying to actually build an AI agent, what agent makes sense to you?

Like it's going to be one of many of those that are that are being available in the market. So, that's sort of like there's a big transition in the market.

And I think it creates complexity as well.

Like, you know, for enterprises, like in terms of how do I keep up with all of this?

Like what are the models I should be using, which ones I should not, which are secure, which are not? So, there's a lot of questions, but but at the same time, a lot of amazing innovation to take advantage of.

And what to pay for, right?

I mean, like you said, this shift that's happened where it's become so much more competitive, and maybe totally commoditized.

Do you need to pay for OpenAI APIs versus, you know, DeepSeek, which is, like efficient right?

Yeah. At the core of that is it feels to me distillation.

Is that right?

Yeah. Yeah. And I think the...

well, let's see the, like, part of it.

It doesn't matter. Like, you know how they got where they got it.

Right. Like, you know what matters to me as a customer is what can I use?

Like, you know, what capabilities am I getting right now? And am I getting them at a lower price? Because remember, that for most enterprises,

price? Because remember, that for most enterprises, like training is not the thing that they do.

Like we just need to use models to actually transform our business processes.

And so if we have a model that's cheaper, that's, you know, as performant and, you know, and I can actually reduce my cost by like 100x.

Well, I'm going to use that.

Has that been sort of the major shift as well, on the application side, is that running AI has become cheaper?

Running AI. So interesting.

I'll tell you a fact about our own company.

Our business is growing at a very, very rapid pace.

We're going to have so much more usage of AI models in our system, but we are modeling our cost to actually not rise at the same pace.

Like, you know, like, I would say that we like, we are anticipating that we'll be able to reduce cost for our end customers by 10x, you know, with all the advances that are going to be coming this year.

But that may mean that they're just using more of it right?

Exactly, yeah. The big debate in the market, can you just going back like a little bit though. Can you,

I know you're primarily concerned with the output, what models and how competitive they are at the end, but can you talk a little bit about distillation for someone who's not familiar with that and how that concept has changed the landscape?

Yeah. So, so distillation as a technique, uh, the way to think about it is, um, you take a large model, and a large model is actually good at doing a lot of different things. Uh,

but for your need, you know, you only have one like you, only you need you need a model that can do one thing really well.

So what you do is you will take a large model and you use that as the, you know, as a as a thing that actually trains this, this new model that you're building, which is going to be as capable as the large one on that, only that one task, but because you've taught the model that only needs to do one thing, it actually is able to simplify itself a lot.

And and so you are able to like, you know, build this new model, which is much lower, like in terms of like, you know, parameters or the cost of inferencing it.

Uh, and it actually solves the problem that you have, which is like, you know, that one task that you wanted to actually go and solve in a very nice way. So, so this is a technique that is actually

way. So, so this is a technique that is actually going to be more and more common.

And in fact, we've been doing this for the last two years as well. But but expect that like, you know, there will be a lot of these purpose built models which are all distillations of a larger model, you know, to to a specific task.

Then what was DeepSeek's breakthrough? Was it the fact that its open source and it used distillation?

The breakthrough, like, you know, from an end-user perspective is, well, I got the model that performs amazing and it's very, very cheap, and it's actually very, very fast as well, like, you know, for inferencing. So, I think and it's actually

inferencing. So, I think and it's actually like DeepSeek is not like a, I would not call it like a typical like, you know, like model distillation process, you know, because it's actually still very broadly, you know, it has broad capabilities. It can

actually do a lot of different things.

There's innovation on top of the distillation.

Yeah. Will others come though?

Did it sort of open up a new playbook that other companies, Chinese or otherwise, are able to use to get right to the frontier and make competitive models?

That trend is actually broader than DeepSeek.

Like, you know, there is a lot of research and innovation that's happening in the industry, like in terms of different techniques on how to train models. So you can expect, like,

models. So you can expect, like, you know, many, many more such advances to come. I mean, this one like,

come. I mean, this one like, you know, captured everybody's attention in a big way. But, I think this is going to be

big way. But, I think this is going to be a normal thing for us. Like, you know, as you know, throughout this year you will see lots and lots of different models like, you know, that are great at certain things and like they can do those things at a fraction of the cost of like the frontier models today.

Help me understand another piece of this, which are the benchmarks and things like humanity's last test, right? These are like math and coding tests.

test, right? These are like math and coding tests.

And, does it open up the question especially, you know, as Glean uses this for other enterprises, the idea of generalization, how do we know that these models aren't just good at passing these specific benchmarks? How do we know that it's going to be able

benchmarks? How do we know that it's going to be able to, you know, solve or analyze or problem solve these sort of unknown tasks that are so unique to an enterprise?

Yeah. So, so that's that's sort of like Glean does, you know, we are actually taking all of this innovation that's happening in the industry, like, you know, like whether the innovation is coming from OpenAI or Anthropic or Google or, you know, Meta or all these open-source models that we have out there. Um, all of them have some capabilities and they can actually solve real world needs. And so what we do at Glean is actually take them

needs. And so what we do at Glean is actually take them all and make them all available to our enterprise customers in a safe and secure way, and actually make it easy for them to make these models work for real world, business-use cases like connecting these models with their knowledge, their data, their information, and their business processes, right?

So now, you know, how like, how do you test, like whether these models are good at doing certain things or not?

In the real world? In the real world.

I mean, I think like we are seeing it like, you know, with our, you know, our application and our, you know, which is primarily about like, knowledge access, like, you know, people are using it very actively in their day-to-day work. Like, you know, whenever you have questions,

work. Like, you know, whenever you have questions, you know, that you need answers for or information that you need to complete your tasks like AI is actually amazing at actually, you know, at that particular category of tasks like, you know, making knowledge more accessible, uh, you know, helping you process, analyze, research, you know, on large amounts of information. So, that's already proven. Like,

information. So, that's already proven. Like,

you know, these models do a great job at at actually analyzing information, making it more accessible to you. But, then if you think about more and more like,

you. But, then if you think about more and more like, you know, like you heard last year, agentic and agents, like, you know, that's the talk of the town.

Like today, like, you know, how businesses are thinking about AI is that, well, I got, like, you know, these business processes, and, you know, they're time consuming.

You know, there's a lot of money that I have to spend, like running these processes.

I want to see if I can build an agent.

So then, you know, you start a process of building an agent to solve a specific business problem.

You do some iterations, you work with a platform like ours or some other.

You actually try different models.

See, like, you know, which one is actually, you know, doing better, you know, for you on that particular task.

You have to tweak. You have to like, you know, it's engineering, like you, you do some work, but ultimately, like, you know, you get there, you actually, you know, bring that automation, like you bring 90% automation or 100% automation in these, you know, processes. So, you're seeing that real

processes. So, you're seeing that real world impact already happen, like, you know, across a large range of problems. So, the models are not like, you know, certainly not like, I won't classify them as being like good at, like you know, giving SAT or just that like, you know, they're actually really techincal.

The ones that you're sort of tweaking and adapting for your customers?

Yeah.

Those are the ones that are better at real world?

No, no, no, all these models are actually good. Like, you

know, like don't don't think of any one of these models that only knows how to take SAT, you know.

Okay, got it.

Generally, these models have broad capabilities and you can actually bring it into a business and make it work for a real business.

Is that because of the reasoning breakthrough?

And then maybe put like the deep research features into context.

That's right. The real power of LLMs have always been reasoning, and some generation, of course, like it can actually use those reasoning capabilities to also write and generate, you know, artifacts for you.

And, that reasoning capabilities are on the rise. Like, you know, you see,

rise. Like, you know, you see, you know, with GPT with like, you know, 01, the level of reasoning capabilities is like, you know, miles ahead of like, you know, the previous generation, um, with 03, like, you know, we're expecting even more 03, you know, is giving the same level of performance as 01 on certain tasks.

So, so that, that the more reasoning that you get because if you think about, you know, any business process, you know, in an enterprise today, you know, there is a, you know, it sort of involves a human, you know, that works with some information and, you know, use some of their reasoning powers and do some work, right? And so, like, you know,

work, right? And so, like, you know, the the reasoning capability, the more you bring from the models, the more complex business processes you'll be able to automate with AI.

Does that mean the data wall is irrelevant now?

Data walls are relevant.

Well, we talked so much about it last year, right? And we thought, okay, maybe the advancements

right? And we thought, okay, maybe the advancements had plateaued. But, it turns out there's just a

had plateaued. But, it turns out there's just a different kind of advancement, right?

Yeah. Yeah. There is, I mean, yeah, there is there is no plateauing on any front, by the way. Like, you know, when it comes to AI and like, you know, the ability for us to leverage AI, you know, in our day-to-day work.

Like, you know, we're just scratching the surface. You know, we talked about this last

surface. You know, we talked about this last year that the event like, you know, even if like, you know, somehow there was a standstill and there was no progress and, model capabilities, like, you still haven't tapped into even 1% of, like, all the capabilities existing models have.

So, there's a lot of work, you know, you're going to see this massive transformation this year, like lots and lots of business processes are going to see AI getting infused in them.

Talk about the OpenAI of it all because their moat, their advantage used to be sort of these pre-training advancements, and the reasoning advancements are really significant, and they're building features on top of it. But, distillation DeepSeek means that sort of it's open season now.

That those advancements are almost commoditized, right? Where does this leave OpenAI?

right? Where does this leave OpenAI?

And, I also wonder how much you're using OpenAI over the last few weeks and what you're doing for your customers?

Yeah. Well, we use OpenAI as well as like these other models from Google and Anthropic and other companies.

Has your usage though with the rise of DeepSeek, has your usage of OpenAI gone down?

No it hasn't. Well, so far, like for Glean right now we've actually evaluated and made we're making DeepSeek available in our platform, but, you know, like it takes time.

Like right now. Like, you know, there's no impact in terms of actual enterprise usage, you know, for our customers.

Like, you know, they're all still using, you know, the same like, you know, frontier models from, from...

Why? Because if DeepSeek offers the same performance, or similar performance at a much cheaper cost, is it only a matter of time to switch over?

Yeah, yeah, it's a matter of time.

You know, there is like, you know, an enterprise, you have to there are a lot of other considerations, like one of them is, of course, security and making sure that like, you know, we're using a robust, you know, model. Like any model that comes in like,

model. Like any model that comes in like, you know, there are possibilities of, you know, prompt injection attacks, and so say like there is a change management process.

Part of it is like, you know, when we are delivering a model to our customer, we need to make sure that it's robust on all fronts.

But second, also like I would say that the cost is not the only, you know, like factor driving, you know, AI adoption and usage today.

Like we're in very early journey.

Like a lot of our customers, they're interested in actually seeing some magic happen. Like it's sort of irrelevant,

happen. Like it's sort of irrelevant, like, you know what the cost of that was because it's not at scale. And so you want to first...

at scale. And so you want to first...

Like for a lot of like new use cases, new agents that you want to build, like the idea like, you know, for in fact, you know, that's what I would recommend to any, you know, any enterprise leader is that, you know, like cost is the second, you know, consideration.

First, like, you know, you have, you know, you want to build an agent that's going to actually transform the business process.

Make it work first, and then figure out like, you know, can you actually optimize it. So if there is actually a clear leader, like in terms of what's the best model, it may be slow, but it's the best, i would start with that first.

But, who is that right now?

Who is that right now?

Well, that depends on the task, like...

Okay. Yeah. But I mean, it just feels like OpenAI had such a hold on that leadership for so long.

Maybe like Gemini and Anthropic, too.

That has changed.

That's changed.

Yeah.

Do you see that mix being a lot different?

I mean...

100%. Yeah, going forward.

I think there's a fundamental shift like, you know, the LLMs, you know, they're going to be many of them and they're all going to get better at different things. And,

you know, as an industry like, you know, it's going to be also like, you know, a commoditized market in that sense, yes.

Do you think that OpenAI is going to become cheaper to for your customers to use?

I think we're already seeing that on the consumer side, right? Right after DeepSeek's RI,

right? Right after DeepSeek's RI, they made 01 free.

That's right. Is that happening on the enterprise side too?

It's going to happen.

We haven't seen immediate pricing changes, but I expect like, you know, like everybody has to be competitive in this market.

And like, you cannot be like, you know, you know, these are real cost differences. This is not like 20%,

differences. This is not like 20%, 30% like these are order of magnitude like reduction in prices. So like, you know, we will see that

prices. So like, you know, we will see that pricing pressure and, like, pricing coming down.

It's really interesting. Last weekend, I think, that Sam Altman did an AMA, Reddit AMA, and he said that OpenAI has been on the wrong side of history when it came to open source, which was just remarkable to me because they've spent the last two years defending it with everything that they have. What does that mean to you?

have. What does that mean to you?

Yeah. Well, you remember that open source always wins. You know, like in tech industry,

wins. You know, like in tech industry, you know, that has been like one truth, you know, of the last three decades is that you cannot beat the momentum.

You know, that a successful open source project is able to actually generate, you know, there is just so much, like, you know, the...

like when you when you build systems in open source, the number of developers who are behind that project, you know, and the amount of innovation that happens there, you know, combined with the fact that, you know, like, that's a technology that's, you know, you can understand it more.

It's like, you know, more efficient, like from a cost perspective.

It's a very, very hard, you know, moment to beat. Like, you know, you have to like, you know, I think the... And like, you know, like when it comes to like, you know, it's not that, you know, everybody gets to always build open source systems. You know, companies like us like, you know, we build a, you know, ready-to-use product for our customers.

But like, you know, that's our mindset too.

Like, you know, under the hood, like, you know, we're going to maximize the use of open source systems that we can like, you know, provided you know that they have the right security, you know, credentials.

Does that mean everything's going to be open source in the future? From ChatGPT to Gemini to Anthropic?

the future? From ChatGPT to Gemini to Anthropic?

Well, I mean, I think the, the..

Well, that's actually a hard, hard question to answer. Like, I don't know.

answer. Like, I don't know.

I think one thing I can tell you is that there's going to be... If you look at usage of models,

be... If you look at usage of models, majority of AI usage is going to be on open source models.

Hm. And I mean, it used to be, when this was sort of a debate, open or closed sources doesn't feel like as much of a debate anymore. It was sort of Llama's going to be the gate... I don't know if gatekeeper is the right way

the gate... I don't know if gatekeeper is the right way because, it's the right word, because it's open. I

mean, who's creating the open source ecosystem?

Is it one player or is it many players?

I think it felt like, you know, open source will be also like 1 or 2 players because, like, the fundamental thing that has changed is do you need billions of dollars to build models.

And that's where, like, now the industry is confident that you don't, and which means that like, you know, the number of players is going to be large. Like there's going to be lots of players who are

large. Like there's going to be lots of players who are going to be innovating and bringing models into open source. Its not going to be,

source. Its not going to be, you know, like one company that is publishing open source models.

Okay, we've only got a few minutes left. I want to make sure I kind of understand this, too, at the enterprise level and the application level, which a lot of people are talking about. You're kind of in the perfect position to see that.

Glean's own financials, right?

I think you guys hit 100 million ARR.

Mhm. Um, has that had anything to do with this shakeup with the cost coming down with all, with distillation, any of these themes?

I would say no. Like, you know, that's a progressive, you know, we hit that like in our last fiscal year. The momentum for us is actually coming

fiscal year. The momentum for us is actually coming from, you know, the fact that, you know, we have a product, that is actually, you know, a very obvious, you know, thing that, you know, our customers want.

Like, you know, think of, like, you know, Glean is... Glean is a better version of ChatGPT,

Glean is... Glean is a better version of ChatGPT, like, for your enterprise.

Like, you know, it's like it can actually be useful to you in your day-to-day work life.

And, I think the, um, people, people actually, like, there's no doubt, like, you know, like every leader in the industry today is focused on AI education.

Like, they know that AI is a big thing.

Like, you know, everything is going to get transformed. You know, their companies is going to

transformed. You know, their companies is going to get transformed and well, to transform the company, you need to make sure that you have the workforce that can actually bring that transformation. And so

there's a lot of focus on AI education.

And when you think about education, like what are the tools? Like how do you actually get people more

tools? Like how do you actually get people more comfortable with AI?

Like, you know, we are a tool which is sort of like, you know, like the most basic one that you would, that you would imagine you want to give to every employee of yours. So,

so that is what has created momentum for us.

I mean, as a user, and I guess that's what people are doing in the enterprise, but just as like a consumer of AI and using the different tools, I really loved R1 because you got the thought process.

Is that something that people in the enterprise value too?

100%. Yeah. Why do you think that is?

Because I think like the magic, you know, that you feel with AI is basically coming from that, you know, human-like reasoning capabilities. Like, you know, the models can

capabilities. Like, you know, the models can think because, you know, like you could always like anything that didn't require thinking where I had to do the thinking, I could actually write a program that could do like the rest of the work.

What machines couldn't do before was, you know, the thinking capabilities and that, like, you know, truly transforms, like, you know, how now you can actually both like what kind of systems you can build, but also like I think this also democratizing, you know, the process of now, like, you know, every business, you know, user, like every employee in the company, they feel like they have the power to create.

Right, and think about the trust side of it, too. You

can see the reasoning and that's so important in the business process.

Arvind, I could talk to you for another hour, and I wish we had another hour, but we'll leave it here. Thank you so much for coming in and talking about

here. Thank you so much for coming in and talking about all these subjects.

Thank you for having me.