Nvidia GPUs as financial products

Nvidia GPUs can be used as collateral to borrow. Financial Times reported that $11 billion of loan was created for these chips.

That’s something too creative for me.

I think it’s generally safe to borrow against assets with a growing value.

Chips, however, are like cars to me, depreciating… with new versions better than the previous one.

Well, it seems you can borrow against your car, but that’s still based on your ability to pay back the loan.

So how does it work?

Maybe it’s actually a loan borrowed against the “service contract” or rental agreement carried by the chips, which makes it more like an asset with “yield”.

But still, this market sounds a bit too arbitrary…

Hard to imagine that the rental income will be stable or rising, as Nvidia chips supply is up to Nvidia and TSMC. That capacity can increase over time.

US big tech maxed out on capex

Assuming buybacks & dividends are “required”, the unallocated cash flows are matching the capex figure already.

This means additional room to add capex is limited.

Put it in another way, free cash flows after shareholder returns are near 0.

For example, Alphabet got $125bn op. cash flow in 2024, paid $53bn in capex, $62bn in buyback and $7bn in dividend. FCF after shareholder returns is only $3bn, or 2.5% of its operating cash flow.

Meta, with 10% unallocated operating cash flow in 2024, will increase its capex by 50% in 2025. If Meta’s op. cash flow grows 10% in 2025, then FCF after shareholder returns in 2025 is only $3bn, or 3% of its operating cash flow.

Microsoft in calendar year 2024, has $10 billion of unallocated free cash flow. But that should be smaller if looking at its fiscal year 2024.

When was Waymo approved?

In 2017, Texas passed a bill to allow driverless cars on the road.

Later that year, Waymo started to bring driverless cars to the road in Texas.

California introduced rules around driverless testing on public roads in Feb 2018.

Waymo won the first driverless permit to test in California in Oct 2018 for ~3 dozen cars.

In 2020, Waymo started to open its fully driverless service to the general public in Phoenix.

In 2024, Waymo offered the service to anyone in SF.


Lots of small steps.

Each state/city can be different. Requirements can be different & definition of “driverless” can be different.

Area can be limited.

Target passengers can be limited.

etc.


Where was Waymo’s technology at?

In 2018, Waymo’s miles per disengagement was 11,154 miles.

In 2023, Waymo’s miles per disengagement was 17,311 miles.

On average, people may drive 10k+ miles per year in the US.

So on average you will only experience one “Disengagement” in a year in 2018, which is a decent rate.


Where is Tesla FSD at?

The latest 12.5 seems to have 1 critical disengagement per 123 miles?

This needs to iterate & improve over time to be fully driverless.

Is Chipotle a robot company that happens to make chipotle?

Not to say these systems are fantastic; but at least you need to experiment first. Investors love stories.

Robots could do fries Chippy robot (by Miso Robotics)

Robots could peel and cut avocados Autocado (by Vebu)

Robots could prepare the bowl/salad (by Hyphen)

 

Problems?

  • cleaning
  • items prepared may not be as good as humans for now
  • lack of customer relationship (I still like it when your local coffee shop people can remember your name)
  • if broke down, hard to replace? unless you have a spare machine nearby… but in every city?
  • less job growth for society
  • etc.

Is LLM still on track of its dream value? Note from podcast

Note from this interesting BG2 podcast: AI Demand / Supply – Models, Agents, the $2T Compute Build Out, Need for More Nuclear & More

Very good discussion. And some notes.

Some would argue that while LLMs are indeed good in coding, chatting etc., that “amazingness” was extrapolated too much. In other words, the gap between pre-ChatGPT and ChatGPT-3 is not gonna continue/replicate. The expectation is too high that the room to beat is diminishing.

Some would question the “moat” of most LLMs. The most advanced ones, the “leading edge” is good, but the rest could be commoditized. Then all those R&D dollars is more of less duplicated and return is low.

GTP-5 is undoubtfully powerful and could be more powerful than people think, but what’s next? What cards haven’t been played or thought about after GPT-5?

Besides technology advancement, there seems to be a problem in regulatory environment in nuclear power in the US, which put a lot of restrictions.

France and China have lowered the nuclear power costs over the years.

 

Future of tech – how AI and blockchain might intersect

AI will be so good at mimicking human.

Blockchain will be the tool to verify origin.


What news, photos, videos should I trust?

I am more likely to believe in those send from my friends whom I trust.

Fake account won’t work if I am already connected with the real account.

The reason I can recognize the “real” account from the “fake” one, even if they look exactly the same, is due to the history (conversations etc.) that happened before, which is what blockchain is about.

First High-NA for Intel

Intel received the world’s first high NA EUV (TWINSCAN EXE:5200) from ASML recently.

The purchase order was made in Jan 2022 – so it took some 2 years.

The machine is said to be 2x as expensive, or $350mn (or more). ASML says it has taken between 10 and 20 orders to date.

If TSMC was to take 10 machines, it will cost $3.5-4bn, or some 13% of its $28-32bn capex target for 2024.

If Intel to budget 35% of its revenue as capex, (say $60bn), it would be ~$21bn, 10 high-NA machines would be 17-19% of that capex budget.

To justify the cost of machine, high-NA “enables building transistors that are about 1.7 times smaller than today, resulting in almost triple the transistor density“. So if density is almost 3x, and assume price per wafer is up by ~45%, then cost per transistor could be halved.

Meanwhile, ASML said 2024 revenue guidance is similar to 2023, which was EUR27.56 billion.

Facebook – Not An Easy Business

Facebook blocked all news content in a Australia on Thursday – users cannot share news links and Facebook Pages of media account are taken down.

This is in response to Australian government’s proposed law, which requires payment deals between media outlets and tech companies over content.

This is also one day after Google stroke a deal with News Corp, the media giant. Under the proposed law, Google will need to pay for news content if they appear in search results.

1/ Why Google and Facebook chose different routes (at least for now)?

I think their ad business are fundamentally different.

Facebook ads is seen on Facebook platforms, but Google ads is seen on both Google products and third-party websites.

Google is enabling third-party advertisers (think about the ads on newspaper’ website) to make money, e.g. AdSense. They are partners, and this network of advertisers is valuable to Google.

Facebook’s ads is sold by getting to know users better and letting users stay on its platforms longer. Traffic is important to Facebook, so news is important as a form of content that users want to see. However, Facebook also thinks it is giving media outlets traffic in return. More important, ads sold by Facebook is not relying on those media outlet.

2/ What content should be on social media?

Instagram is in a purer form of social media, so does Twitter. They are usually gravitating towards certain types of contents. On the other hand, products such as Facebook’ main app are aggregating all kinds of “feeds” as long as they can drive traffic.

I think the two types are both here to stay.

Another related issue is how to regulate contents, which has been an increasingly important issue in the US and globally.

“Regulate more” or “regulate less”?

I think either way more regulatory interventions (government) is most likely inevitable.

If platforms regulate less, regulators may think Section 230 is providing to much protection and platforms are not doing enough for their social responsibility.

If platforms regulate more, regulators might think they have too much power, which is also risky. And as they moderate more, it costs more and they may be challenged more often on their decisions.

“Public square” is not easy. “Digital living room” is where Facebook may find more flexibility in contents.

Paying for news might be one of the solutions to navigate some content risk, e.g. fake news, misinformation. However, fake news or misinformation might be the traffic driver that Facebook values.

Kuaishou Valuation

So the competitor of Douyin (TikTok) in China, Kuaishou (HKEX: 1024) just went IPO this week, now valued at ~$160 billion.

It’s a well-known app in China – with the current market sentiment, hypes around video-based social platforms, I should say I am not surprised about the valuation.

Here I provide one way to look at Kuaishou’s valuation:

Q3 revenue is ~$2.4 billion:

1/ ~50% comes from live-streaming (virtual gifts).

2/ The rest is from ads and e-commerce. I categorize them as “good” revenues that are fast growing (+200% yoy for ads) and stable.

For the first part, virtual gifts, we can use Huya (most revenue is from live-streaming) as a comp – about 3x annualized revenue.

For the second part, we can compare it with Snap, which trades at 25x annualized revenue. Kuaishou’s ads business (~$900 million in Q3) is at the same scale as Snap and grows faster – so some can argue to use 30x.

Therefore,

Kuaishou = virtual gifts business x 3 + ads & other business x 30

As virtual gifts is 50% now, we are talk about 16.5x revenue as a whole.

Annualize it: $2.4 billion x 4 x 16.5 = $158.4 billion


Virtual gifts business is debatable – while it’s 50% of revenue, it accounts for less than 20% of Kuaishou’s value if we use the above framework. So change it from 3x to 2x or 1x won’t affect much actually.

Clubhouse Revenue (?)

Clubhouse is a $1 billion business with 2 million weekly active users (as of Jan 2021).

Despite the exploding user count, it appears to have zero revenue now.

Clubhouse mentioned three business model:  tipping, tickets and subscriptions (therefore, no ads – good!).

1/ tipping – it’s an established model in Asia but don’t think it will work well in the US. If rolled out, it will be on a smaller scale and shouldn’t be the main source of revenue, unless it incorporates cryptocurrencies like Dogecoin!

2/ tickets – I think this can work but it’s not sexy. It makes the app more “administrative” and ordinary. Also, eventbrite works well with some 2.5%+ fees.

3/ subscriptions – this sounds interesting and may leapfrog podcast subscription model. Another layer of curation can be added, which Clubhouse might have been working on).

Clubhouse may also think about adding features to be more relevant on a daily basis, than monetizing the app right now.

Spotify management should be worried.

Social medias should be worried too.