The Perfect Short

It should come as no surprise…

The latest developments surrounding China-based hedge fund High-Flyer Capital Management (HFCM) – and its release of its DeepSeek AI models – have become even more interesting.

For any subscribers who missed yesterday’s Bleeding Edge – Did the Leaders in AI Get it All Wrong, it’s a must-read before reading today’s issue.

As the industry continues to grapple in real-time with the implications of DeepSeek’s high-performance large language models (LLMs) – roughly on par with OpenAI and other prevailing U.S.-based models…

HFCM just released another generative artificial intelligence model.

Janus-Pro-7B.

Another DeepSeek “Best-in-Class”

Janus-Pro-7B is a multimodal, generative AI model capable of text-to-image generation.

And, in line with the performance of DeepSeek-V3 and DeepSeek-R1, the performance appears to be equally impressive.

Janus-Pro-7B Performance | Source: DeepSeek

Worth noting in particular is the chart above on the right, which shows Janus-Pro-7B (in blue hashed) compared to OpenAI’s DALL-E 3 (in light gray) – OpenAI’s most advanced multimodal text-to-image AI integrated into ChatGPT now.

According to DeepSeek, Janus-Pro-7B is positioned as best-in-class compared to the others on the charts.

DeepSeek also took time to highlight the improvements it has made in its text-to-image generation compared to Janus, which is the predecessor to Janus-Pro-7B.

Improvements in Janus-Pro-7B compared to Janus | Source: DeepSeek

It’s very easy to see the dramatic differences between the images on the left compared to the images on the right.

Images on the left clearly look like distorted computer-generated images… and the ones on the right are photorealistic, so much so that most would see them as real-life images.

The media has largely been sucked into the headlines of these latest DeepSeek releases.

On the surface they are impressive, especially coming from a small company in Hangzhou, China.

But the last 24 hours have surfaced – and confirmed – much of what I wrote about in yesterday’s Bleeding Edge.

There’s much more to the story than the screeching headlines would have us believe.

The *Asterisk to the DeepSeek Story

Here’s what the details have now revealed about DeepSeek:

• The actual costs to develop the technology are far more than the quoted sticker price of $5.5 million. That amount may be the compute cost to train the DeepSeek-R1 model but is not inclusive of all the prior research, development, and computing costs to get to this stage.
• For any company or individual that connects to DeepSeek’s APIs, the terms of service provide that customer data is sent back to China. Most companies are choosing not to connect to DeepSeek’s APIs for this reason. It creates a clear and present security risk.
• No analysis has been done yet if data is sent back to China when DeepSeek-R1 or DeepSeek-V3 are run “independently.”
• The software is not actually open-source. Only the model weights are open-sourced, allowing the user some level of fine-tuning of the model. The training data and the full software code are not open-sourced and therefore cannot be verified for security and architecture.
• DeepSeek benefited from existing U.S.-developed models.

These realities evoked a succinct response from a managing partner of Thrive Capital, an investor in OpenAI:

Source: X @JoshuaKushner

Despite these glaring truths, even the Wall Street Journal fell into the trap, with an article like “How China’s DeepSeek Outsmarted America.”

It’s crazy for the WSJ to write something like that without a single acknowledgment of the facts listed above.

But even more surprising is that it missed the elephant in the room…

It’s so easy to get excited about the technology, the advancements, and the possibility of doing things cheaper and more efficiently than before. So much so that it’s easy to forget the company behind DeepSeek…

… a hedge fund – High-Flyer Capital Management.

What business does a hedge fund have for developing the technology to power a foundational large-language model?

Please stick with me, this is about to get even more interesting…

What My Gut Is Telling Me

Short-selling hedge funds all play the same game.

Once they have determined that a company is way overvalued or cooking the books, they build a massive short position – quietly… over a number of weeks or months.

It takes time to build the desired positions and aggregate enough shares to sell short. And the preference is to build those positions through intermediaries, using dark pools so that others can’t see what you’re doing.

Once the fund has its desired short position, then it goes to CNBC, the WSJ, Bloomberg, etc. to talk up its short idea – all in an effort to get institutional capital to panic-sell their positions.

Notable short-sellers that have done this in the past are Jim Chanos of Kynikos Associations, Michael Burry of Scion Asset Management, David Einhorn of Greenlight Capital, Bill Ackman of Pershing Square, Fahmi Quadir of Safkhet Capital, as well as short-selling research firms like Hindenburg Research and Citron Research.

Is it that much of a stretch to believe that a hedge fund – in this case, HFCM – may have built up a massive short position in NVIDIA… in the months leading up to the timed release of DeepSeek-R1, which it controls, knowing the impact it would have on NVIDIA’s stock?

When we think about it, we realize it’s the perfect short:

• HFCM already has deep expertise in working with AI and NVIDIA GPUs for its own hedge fund.
• It sees that it can develop its own large language model, largely leveraging existing technology available as open-source software (i.e. it can develop cheaply).
• Position the software publicly as having been developed cheaply, and capable of running the model at a fraction of the cost of prevailing Western models.
• Cause mass panic in the market, suggesting that the industry doesn’t need all these GPUs and that they can just change their architecture.
• NVIDIA’s share price plummets on these developments.
• HFCM closes out the short position, taking a massive profit off the table (hundreds of millions?).

NVIDIA is up more than 7% today. Could part of that be a hedge fund covering their short position and taking their money off the table? Absolutely.

Do I have any hard proof? No, I’m just speculating. This is my gut speaking from decades of experience. It feels obvious to me as a distinct possibility.

And the short would come with the added benefits that I shared yesterday, in terms of China hoping to slow U.S. development down so it can catch up… and position another backdoor into U.S. consumer data in the event TikTok is shut down.

But putting all of this fun and clever speculation aside, another gut reaction has now been made very clear…

A Shot of Adrenaline

The DeepSeek developments have lit a fire royale under the U.S. development of artificial intelligence.

From the White House in Washington, D.C., to the venture capitalists on Sand Hill Road, everyone just took a shot of adrenaline right in the heart.

If it was a sprint before, it’s a mad dash now, and they will not slow down for anything.

In fact, they are going to speed up.

Just have a look at how the tone has changed from Sam Altman, founder and CEO of OpenAI. Here’s a post from January 20, pre-DeepSeek announcement.

Source: X @sama

It was a clear attempt to de-hype and reduce expectations about what OpenAI has been working on and what’s around the corner (i.e. he didn’t want to tip off his competition as to how far along OpenAI was with its development).

Here’s yesterday’s post. Note Altman’s shift in tone:

Source: X @sama

We “believe more compute is more important now than ever before to succeed at our mission.”

“The world is going to want to use a LOT of AI and really be quite amazed by the next-gen models coming.”

And the kicker…

“look forward to bringing you all AGI and beyond.”

AGI – artificial general intelligence.

We can imagine what happened to prompt Altman to step up and tell the truth: That OpenAI is actually working on AGI.

A bunch of OpenAI’s investors hammered Altman after the DeepSeek news coming out with major questions about the competitive landscape, the architecture, and whether or not the massive Project Stargate-like data centers are necessary.

I’ll cut right to the chase…

Short answer: If anyone wants to build a foundational model, an AGI, you’re damn right we need the computational horsepower.

And once AGIs (plural) have been built and are being used by billions of people around the world daily, are we going to need more data centers or fewer?

My Bleeding Edge subscribers are curious and smart, and they know the answer to that question.

Jeff

P.S. You can go right here if you’d like to learn more about how we utilize our own artificial intelligence – a neural network called our Deep Access AI – to help us identify stocks with a high probability of declining, and how to develop trades off of these signals…