Fountainheadinvesting

Fountainhead Investing

  • Objective Analysis: Research On High Quality Companies With Sustainable Moats
  • Tech Focused: 60% Allocated To AI, Semiconductors, Technology

5 Star Tech Analyst Focused On Excellent Companies With Sustainable Moats

Tag: Artificial Intelligence

Categories
Ad Tech AI Cloud Service Providers Industry Stocks

Amazon Is A Good Bargain At $202

AWS

I’ve been adding Amazon (AMZN) to my portfolio in the past week; it is a bargain at $202, having dropped almost 20% from its high of $242.

Amazon has 4 businesses.

Amazon Web Services

AWS is a cloud services behemoth and market leader with $ 108 Bn in 2024 sales, and still growing at 19%. That is remarkable growth for a market leader of that size with two other 800-pound Gorillas, Alphabet and Microsoft, chasing it. It generated operating profits of $39 Bn last year, a growth of 66% with an operating profit margin of 37%. This is Amazon’s most profitable segment and the growth engine, which powers everything.

Advertising

Amazon includes its advertising revenues in the online retail sales segment, but its advertising revenue last year was estimated between $ 56 Bn to $ 64 Bn in 2024, growing around 20% a year. This is also a high operating margin business, generating over 20% in operating profits.

Prime Subscriptions

Amazon doesn’t disclose its Prime subscriber numbers, but we estimate about 200Mn subscribers, including 180Mn in the US in 2024, generating over $40Bn in revenue.

This is another sustainable, sticky, and high-margin business, I’d value it at about 9x sales or $360 Bn.

I used a 9-10x multiple for the high-growth, high-profit margin, and sustainable businesses.

Online and physical retail sales in the US and abroad

These include third-party sales. Physical sales revenues are minuscule compared to total retail sales; loss leaders to expand reach and for analytic purposes, including in the online retail business. Amazon had a whopping $ 431 Bn in 2024. While online domestic and international sales are a drag, growing slower in single digits, they’re not significantly slower than Walmart’s sales growth and margins.

Amazon Segment Sales: Sources Amazon

Based on the Sum of the Parts schedule above, we’re getting the online and physical retail operations of $431 Bn at a market cap of just $170 Bn. The multiple of 0.4 is much lower than Walmart’s multiple of 0.74, or 40%.

Amazon has been spending heavily on Capex for AI to gear AWS and expand its web service offerings. In this arms race, they are scheduled to spend $100 Bn in 2025 to maintain and possibly expand their leadership.

We haven’t even valued all their investments and partnerships under AI development. That can be very valuable in the future.

I’d continue to buy the stock on declines.

Categories
AI Industry Semiconductors Stocks

Nvidia GTC Keynote – CEO Jensen Huang

Nvidia Agentic AI

Quick Key Takeaways: Worth every minute. I own shares and will add on declines.

Incredible product road map:

Blackwells are in full production and Blackwell NV72 is expected in 2H2025. Some may quibble about a “delay, ” as investors expected Q1 or Q2 of strong sales from NV36 and NV72, but it hardly makes a difference in the long run. In the worst-case scenario, the stock could drop 10-15%, but that should attract buying unless other macroeconomic uncertainties cause a continuous slump in the market/economy. I would back up the truck at $100.

Rubin, which is the next series of GPU systems, will be available in the second half of 2026 – again a massive leap in performance.

Nvidia (NVDA) has a 1-year upgrade cadence,

a)Nobody else has that

b) It’s across the board, GPUs, Racks, Networking, Storage, and Packaging the whole ecosystem of partners.

Nvidia’s market leadership is going to last a while – That is my main investing thesis, and I can withstand the short-term bumps.

Cost Analysis: I’m glad Jensen spoke about this in more detail and what stood out for me was a clear-cut analysis of reducing variable costs as their GPU systems get smarter and more efficient to bring total costs (TCO) down. Generating tokens to answer queries is horrendously expensive for Large Language Models, like ChatGPT and it was a black hole.

I expect costs to come down significantly, making the business model viable. Customers are expected to pay up to $3Mn for the Blackwell NV72, and it has to become profitable for them.

Omniverse, Cosmos, and Robotics, – are other focus areas to go beyond data centers. Nvidia needs other target markets, industrials, factories, automakers, and oil and gas companies to embrace AI and therefore use of their GPUs, to reduce their dependence on hyperscalers, and Jensen spent a lot of time on them. He also emphasized enterprise software partnerships, for AI, and gaining full acceptance as a ubiquitous product and making extra revenues. For Nvisia’s vision of alternate intelligent computing, we have to see more Palantirs, Service Nows, and AppLovins. In my opinion, Agentic AI will make serious inroads in 12-15 months.

I’ll add more detail in another note, once I parse the transcript in detail.

Bottom line – this is going to be an incredible journey and we’re just at the beginning; Sure it’s going to be a bumpy ride, and given the macro environment, it would be prudent to manage risks by waiting for the right entry, and taking profits when overvalued or overbought, or if you have the expertise, using other hedging mechanisms. I’m sold on Jensen’s idea of a new paradigm of accelerated and intelligent computing based on GPUs and agents. Nvidia is best positioned to take full advantage of it.

Categories
AI Cloud Service Providers Industry Semiconductors Stocks

Credo Technology (CRDO) $46 Is A Great Pick And Shovels Play On AI

Credo Technology Is Essential For Data Center GPUs

While all eyes and ears are on tariff uncertainties and geopolitical risks, we remain focused on finding good investments for the long term – tuning out the drama and volatility.

Excellent Q3-FY2025 results

Credo Technologies (CRDO) supplies high-quality Active Electric Cables (AECs) to data centers, counting on Amazon, Microsoft, and other hyperscalers as its biggest customers. The stock dropped 14% today to $46.75 in spite of excellent Q3-FY2025 results with a 154% increase in sales to $135Mn Vs $120Mn expected and a sizable improvement in gross and operating margins, which is unusual when you’re ramping up production for a customer like Amazon.

Revenue guidance for the next quarter was even more impressive at 162% growth to a midpoint of $160Mn. For the full year ending in April 2025, Credo is expected to grow revenues to $427Mn – a whopping 121% increase, over the previous year.

Good pick and shovels play in data center and AI

Credo is a pick and shovels AI/GPU/Data center play as data centers ramp up all over the world for accelerated computing. Its key products are essentially AEC replacements for optical cables — a play on back-end networking of high, and reliable bandwidth for data center GPUs and GPU systems like the Nvidia Blackwell N36 and N72, which are expected to start ramping up in the 2nd quarter of 2025.

Data center equipment suppliers have become very crucial parts of the AI/GPU supply chain, and Credo’s results certainly speak volumes of their capacity to scale and scale profitably, which is even more admirable.

Its founders are from Marvell (MRVL), there is a fair amount of credibility and experience.

They are general purpose and custom silicon agnostic, which is good because get business from Nvidia and from ASIC players like Amazon and Google.

The business is also GAAP breaking even in FY2025, another exception for such a small company.

Credo had gross GAAP margins of 63.6%, and GAAP operating margins of 20% and a stunning Adjusted Operating Margin of 31.4%, which is astonishing for a fledgling 400Mn operation with Amazon as its main customer.

Key Risks 

Customer concentration – not likely to change soon, the nature of the industry currently needs high volume from hyperscalers.

AEC cables will become a commodity after 3-5 years, so they’ll need to maintain their growth without dropping prices.

Valuation

Credo’s valuation is not expensive at 11x sales as the revenue growth is easily going to surpass 60% in FY2026 and 30% in FY2027, after growing 120% in FY2025. The P/S to growth ratio drops to a low of 0.2 with such high growth. Furthermore, it has an operating profit margin of 20% easily adding to more than the rule of 40, or 60+20 = 80.

The drop today was ostensibly because of customer concentration – Amazon 68%. But analysts and investors should have known this; I believe the correction is overdone and Credo should resume its upward march again. I bought at 45.75 today, the stock is down almost 50% from its all-time high of $86.69, but still up 187% in the past year.

I’m targeting a return of 24% per year or double in 3.

Categories
AI Cloud Service Providers Industry Semiconductors Stocks

Nebius (NBIS) $45 Has Long-Term Potential But May Be Priced To Perfection

Source: The Next Web


While Nebius has shot up 135% in the past year and is approaching fever pitch as a speculative AI infrastructure investment, it does have long-term potential to justify buying on declines.
Nebius was carved out of the Yandex group, an erstwhile Russian company, known as the Yahoo of Russia. After sanctions due to the Ukraine war and the resulting spinoff, this is a European company with US operations with little or no Russian exposure or additional geo-political risks.
Nvidia has a 0.3% stake in the company, and a strategic partnership to expand AI infrastructure to Small and Medium businesses beyond hyperscalers.
Nebius has five revenue segments – Data Center, Toloka, Triple Ten, Clickhouse and AVRide.

I want to focus on the main data center segment in this article.
Datacenter
The best and most strategic segment is the data center, and the key reason to invest in the company is to take full advantage of AI needs beyond the hyperscalers. I expect at least 100% annual revenue growth in the next two years from the data center, slowing down to 50% in year 3.
Nebius is going all out in creating enough capacity for demand in the next two to three years.
It launched its first data center in the US, in Kansas City to start operations in Q1 2025 with an initial capacity of 5 MW, scalable up to 40 MW.
Further expansion plans – most likely all of that is Nvidia’s B200 GPUs.
• Finland 25MW to 75MB by late 2025 or early 2026.
• France – 5MW Launching in November 2025.
• Kansas City – Second facility with 5MW to expand to 40MW.
• One to two further greenfield data centers in Europe.
Datacenter offerings: Either as computing power or GPU rentals, or the more specialized PaaS (Platform As A Service) with its AI studio, which gives customers choices of OpenAI or DeepSeek models among others. It is priced based on usage and token generation to cater to medium-sized, smaller, and/or specialized domain-specific customers.
Fragmentation likely: As the AI data center industry progresses, I believe, Inferencing and modeling requirements will be fragmented and domain-specific. The DeepSeek software and modeling workarounds do suggest this market could easily be targeted by customized requirements, where brute computing power as the norm will morph into specialized or customized requirements. In which case while customers could contract for larger GPU training clusters, they would also look for cheaper inference solutions, which rely on software enhancements. This is likely to happen over time and may well work to Nebius’ advantage since they want to go beyond pure GPU rentals and provide a full stack – this could be both a challenge and an opportunity and it would be crucial to get more visibility in Nebius’ offerings or services as the year progresses specially compared to competition like CoreWeave.
Lowering training costs: While there remains a huge cloud about what DeepSeek did spend on training, and even as 2025 seems to be secure because of the large Capex of $320 Bn committed by the hyperscalers, it would be remiss to not acknowledge that the trend would seek lower training costs as well – in which case a) data center computing power will be at risk and b) pricing could be the main differentiator. As of now, Goldman Sachs is projecting data center demand to exceed supply by about 2:1, and the gap is unlikely to be filled even with rapid deployment through 2026.
Spend more to earn more: Most of the forecasted growth is based on Capex possibly over $2.5Bn in 2025 with an additional $2.5Bn to $3Bn in 2026. Currently, Nebius is well capitalized with about $2.9Bn in cash, but if data centers don’t generate enough cash, there could be dilution to raise more capital or the sale of stakes in their 4 other businesses. This is very likely to happen in 2026.

Negatives and challenges

Provide value to customers beyond brute computing power: Over time data center rentals will get commoditized and become price-sensitive. The DeepSeek modeling workarounds do suggest that brute computing power will morph into smaller specialized or customized requirements. This could also work to Nebius’ advantage, since they can provide a full stack, i.e. –a challenge and an opportunity.
Pricing could be a challenge: The trend towards seeking lower training costs should continue – as of now Goldman Sachs is projecting data center demand to exceed supply by about 2:1, and the gap is unlikely to be filled even with rapid deployment through 2026, but Nebius needs to stay on top of it, to ensure they generate enough cash to continue spending on growth.
High Capex Needs: Most of the forecasted growth is based on Capex possibly $2.5Bn to $3Bn in 2025 and 2029 each – currently Nebius is well capitalized with over $2.9Bn in cash, but investors will need to be patient with this outlay first and prepared for dilution.

Valuation:

Nebius had forecasted to reach an ARR (Annual Recurring Revenue) Run Rate of – $750Mn to $1Bn for 2025, which is based on about $60-80Mn of ARR in Dec 2025 times 12. It’s not the ARR in February. It would grow from the current $300 Mn to $ 875 Mn by the end of 2025. Normally annual revenues are much lower than ARR – a lot of ARR is deferred revenue because the ARR includes contracted revenues, which are then pro-rated for the year. An ARR of $875Mn at the midpoint could imply 2025 revenue of between $400 and $500Mn. (This is still about 3x the estimated 2024 revenue of $137Mn – so there is tremendous growth. (But at this stage a lot of estimates!)
At a market cap of $10.4Bn, we’re looking at over 20x to 25x, 2025 revenue, so the price may have gotten ahead of itself.
This is a thinly traded company with rampant speculation, and I think the best move would be to sell 25% to 50% before earnings, should the quarterly results and forecasts disappoint. I’m already making a decent profit in a short time and keep the rest for the long term. Nebius reports pre-market on Thursday 20th, Feb.
I would like to see more visibility before committing to invest more.

Competition

CoreWeave (which is private) and also an Nvidia strategic partner had estimated revenue of $2.4Bn in 2024, and with the addition of 9 new data centers to 23 very likely to have around $8Bn of sales in 2025.
CoreWeave was last valued at around $23Bn but is targeting an IPO valuation of $35Bn thus giving it an estimated sales multiple of anywhere between just 4-9x for 2025, way below Nebius.
Even if we assume that the other businesses contribute an additional 25% or $125Mn in 2025 revenues we’re still valuing Nebius much higher than CoreWeave – a larger and more established competitor with Nvidia as a partner, and Microsoft as a customer.

That makes me wary; I’d be happy if my sales estimates are too low, but if they are not, then I would rather wait for dips.

Categories
Industry Stocks Technology

Confluent’s Excellent Quarter Is A Major Inflection Point

Confluent Data Streaming

02/11/2025

Confluent (CFLT) $37 – Still worth buying.

I’ve owned it for over two years but will pyramid (add smaller quantities on a large base) it further.

Why is this company still worth investing in after a 20% post-earning bump?

Four important catalysts

Databricks partnership: The partnership with Databricka, which is much better known and valued increases brand awareness and opens a lot of new opportunities and doors.

This could accelerate growth from the current 22-23%.

Strong customer base: 90% of its revenues are coming from 100K + ARR clients.

The $1Mn+cohort saw the highest growth, and Confluent managed a net ARR of 117%, indicating strong upselling.

A changing data processing market: The entire batch processing model could be up for grabs – customers moving at the speed of light and willing to pay for the latest technology could be a huge TAM. 

This is a paradigm shift, which Confluent has been trying to build into for a decade. 2025 might be that inflection year, with all the AI build-outs and use cases that are likely to need live processing – Confluent is the leader in that field. To be sure it’s not going to throw data processing models into obsolescence, why would you spend money on data that doesn’t need to be processed in real-time, but could take a large chunk of that market?

Snowflake acquiring RedPanda: Snowflake is reportedly trying to buy streaming competitor RedPanda for about 40x sales: While it’s not an obvious comparison, Red Panda is supposedly less than 10% of Confluent’s revenues but growing at 200-300%. But it’s the synergy with the larger data provider that’s getting it a massive price tag – Snowflake would love to have this arrow in its quiver of data tools.

Confluent is best positioned to take advantage of the possible shift from batch processing to processing in data streaming; its founders invented Apache Kafka, the open-source model for data streaming. And while its own invention is available for free – managing and maintaining it at scale needs the paid version. Over the years with the focus on Confluent Cloud, Confluent gets 90% of its $1Bn revenue from customers over $100K in annual revenue. 

Confluent has the cash the tech chops and the focus – sure Apache Kafka is open source and many cloud service providers like AWS and Microsoft also provide enough competition, but no one has the product breadth that Confluent does.

I would not be surprised if Confluent’s multiple expands from the current 8x sales after this earnings call.

Here are the details of the December 2024, 4th quarter earnings:

  • Q4 Non-GAAP EPS of $0.09 beat by $0.03.
  • Revenue of $261.2Mn (+22.5% Y/Y) beat by $4.32Mn.
  • Q4 subscription revenue of $251Mn up 24% YoY
  • Confluent Cloud revenue of $138Mn up 38% YoY
  • 2024 subscription revenue of $922Mn up 26% YoY
  • Confluent Cloud revenue of $492Mn up 41%YoY
  • 1,381 customers with $100,000 or greater in ARR, up 12% YoY.
  • 194 customers with $1Mn or greater in ARR, 23% YoY.

Financial Outlook

Q1 2025 OutlookFY 2025 Outlook
Subscription Revenue$253-$254 million$1.117-$1.121 billion
Non-GAAP Operating Margin~3%~6%
Non-GAAP Net Income Per Diluted Share$0.06-$0.07 vs. consensus of $0.06~$0.35 vs. consensus of $0.35

Categories
AI Industry Semiconductors Stocks

ASML – An Excellent Company That’s Still A Bargain

ASML- EUV

The Monopoly

As the source of all things AI and related, ASML is (and has been for the past decade) the monopoly for EUV lithography machines that power the most advanced GPUs from Nvidia and others. There’s no other manufacturer that can do this at scale.

Around October 2024, on their earnings call, they disappointed the market with a 10-15% lower than forecast revenue for 2025-2026, as one of their customers (very likely Intel) did not place orders for a large order of EUV machines as expected. Intel’s troubles are well known and this order is unlikely to come back. They also feared export controls to China and/or weakness in Chinese demand after 3-4 years of rapid growth.

I bought and recommended buying on 10/27/2024 at $690, with the following comments

Sure it could stay sluggish, range-bound, or fall till there’s some improvement in bookings, export controls to China, etc. Perhaps, that may not even happen for a while.

I think that’s an acceptable risk, now I’m getting a monopoly at a 37% drop from its 52-week high of $1,110, still growing revenue at 12% and EPS at 22%, selling for 8x sales and 25x earnings.

With TSM’s results, we saw how strong AI semiconductor demand still is and there was absolutely no let-up in their guidance.

A monopoly for AI chip production – an essential cog, without which AI is not possible – is definitely worth the risk

Fast Forward to the next quarter, the dynamic is much better and the price hasn’t shot beyond affordable.  

Bottom line: A must-have, it’s always going to be priced at a premium given its monopoly status and the strength of the AI market, so returns are likely not going to be like a fast grower tech but I’m confident of getting 14-16% annualized return in the next 5-10 years.

ASML’s Q4-2024 results on 01/29/2025 were excellent:

ASML beats expectations as bookings soared.

The EUV, machines leader grew Q4 revenues 28% YoYr to €9.26B, and 24% QoQ, beating estimates.

Bookings: ASML’s Q4 bookings came in huge at €7.09B, way ahead of estimates of €3.53B., with net new adds of €3B.

On the earnings call, CEO Christophe Fouquet had this to say about AI and sales to China:

AI is the clear driver. I think we started to see that last year. In fact, at this point, we really believe that AI is creating a shift in the market and we have seen customers benefiting from it very strongly. Others maybe a bit less.

We had a lot of discussion about China in 2023-2024 because our revenue in China was extremely high. We have explained that this was caused by the fact that we are still working on some backlog created in 2022, when our capacity was not big enough to fulfil the whole market. 2025 will be a year where we see China going back to a more normal ratio in our business. We are going to see numbers people used to see before 2023.

USA led sales with with 28% share in the fourth quarter of 2024, edging China’s 27% of about €7.12Bn

Challenges remain as the AI arms race gets hotter:

ASML has not been able to sell its EUV machines to China because of U.S.-led export curbs to restrict China from getting advanced lithography equipment to manufacture cutting-edge chips like the H100s from Nvidia, or the new generation Blackwells.

From 2025, ASML will provide a backlog of orders on an annual basis instead of bookings to more accurately reflect its business.

Guidance: 2025 total net sales remain the same, between €30B and €35B. Q1-2025 is slightly higher with total net sales to be between €7.5B and €8.0B versus consensus of €7.24B.

ASML remains an excellent opportunity and I plan to add it on declines.

Categories
AI Enterprise Software Industry Market Outlook Stocks

AI And The Multiplier Effect From Software

The AI Software Revolution

02/11/2025

The Software Multiplier Effect: An interesting note from Wedbush’s Dan Ives on Artificial Intelligence, who believes that software AI players will likely get 8 times the revenue of hardware sellers. I.e., a multiplier effect of 8:1 from software.

He is directionally right, and I do agree with him about the multiplier effect of software, services, and platforms on top of hardware sales. I had done a primary study several years ago with companies like Oracle, IBM, and Salesforce among others, and we saw similar feedback of about 6 to 1 for software spend to hardware spend, over time. People naturally cost more.

Nonetheless, regardless of whether it is 6 to 1 or 8 to 1, both numbers are huge and extremely likely in my opinion in the next 5 to 10 years and Palantir’s (PLTR) Dec quarter earnings hit it out of the park.

Dan Ives said:

Palantir Technologies (NASDAQ:PLTR) and Salesforce (NYSE:CRM) remain the two best software plays on the AI Revolution for 2025.

The firm also recommended other software vendors such as Oracle (ORCL) IBM (IBM), Innodata (INOD) Snowflake (SNOW), MongoDB (MDB), Elastic (ESTC), and Pegasystems (PEGA) enjoying the AI spoils.

Analysts led by Daniel Ives said:

Palantir has been a major focus during the AI Revolution with expanding use cases for its marquee products leading to a larger partner ecosystem with rapidly rising demand across the landscape for enterprise-scale and enterprise-ready generative AI.

Major Growth Expected: The analysts added that this will be a major growth driver for the U.S. Commercial business over the next 12 to 18 months as more enterprises take the AI path with Palantir. They believe “Palantir has a credible path to morph into the next Oracle over the coming decade” with Artificial Intelligence Platform, or AIP, leading the way.

Wedbush’s feedback about budget allocations is very helpful and even if one discounted Dan Ives’ perpetual optimism and bullishness by some, it’s a great indicator that this will be a favored sector in 2025-2028.

Ives and his team have been tracking several large companies that are or are planning to use AI path in 2025 to gauge enterprise AI spending, use cases, and which vendors are separating from the pack in the AI Revolution.

The numbers are gratifying:

Analysts expect that AI now consists of about 10% of many IT budgets for 2025 they are tracking and in some cases up to 15%, as many chief information officers, or CIOs, have accelerated their AI strategy over the next six to nine months as monetization of this key theme is starting to become a reality across many industries.

“While the first steps in AI deployments are around Nvidia (NVDA) chips and the cloud stalwarts, importantly we estimate that for every $1 spent on Nvidia, there is an $8-$10 multiplier across the rest of the tech ecosystem,” said Ives and his team.

What’s more important?

Analysts noted that about 70% of customers they have talked to have accelerated their AI budget dollars and initiatives over the last six months. The analysts added that herein is the huge spending that is now going on in the tech world, with $2T of AI capital expenditure over the next three years fueling this generational tech spending wave.

Hyperscalers indicated supreme confidence in their AI strategy committing in excess of $300Bn in Capex for 2025, which is historic. Amazon’s CEO Any Jassy was categorical in stating AWS doesn’t spend till they’re certain of demand.

Ives had this to add, underscoring Amazon’s confidence.

In addition, Ives and his team said that they are seeing many IT departments focused on foundational hyperscale deployments for AI around Microsoft (MSFT) Amazon (AMZN), and Google (GOOG) (GOOGL) with a focus on software-driven use cases currently underway.

“The AI Software era is now here in our view,” said Ives and his team. Wedbush’s team strongly believes that the broader software space will expand the AI revolution further, cementing what I saw at the CES last month. There is so much computing power available and so many possibilities of use cases exploding that this space could see a major inflection point in 2025-2026.

Large language models, or LLM, and the adoption of generative AI should be a major catalyst for the software sector.

Categories
Cloud Service Providers Industry Semiconductors Stocks Technology

Marvell’s Investment Case Got Stronger With Hyperscaler Capex

Marvell Custom Silicon

Marvell Technology (MRVL) $114

I missed buying this in the low 90s, waiting to see if their transformation to an AI chip company was complete. Having a cyclical past, with non-performing business segments made me hesitate, besides far too many promises have been made in the AI space only for investors to be disappointed.

Marvel has been walking the talk, Q3 results in Dec 2024 were exemplary, and guidance even better.

Hyperscaler demand

With a planned Capex of $105Bn for 2025, Amazon confirmed on their earnings call that the focus will continue on custom silicon and inferencing. Amazon and Marvell have a five-year, “multi-generational” agreement for Marvell to provide Amazon Web Services with the Trainium and Inferentia chips and other data center equipment. Since the deal is “multi-generational,” Marvell will continue to supply the released Trainium2 5nm (Trn2) while also supplying the newly-announced Trainium3 (Trn3) on the 3nm process node expected to ship at the end of 2025. Amazon is an investor in Anthropic with plans to build a supercomputing system with “hundreds of thousands” of Trainium2 chips called Project Rainier. The DeepSeek aftermath does suggest a further democratization of AI, as inference starts gaining prominence from 2026.

Critically, like other hyperscalers Microsoft, Meta, and Alphabet, Amazon announced a high Capex (Capital Expenditure) plan of $105Bn for 2025, 27% higher than 2024, which itself was 57% higher than the previous year, for AI cloud and datacenter buildout. It was the last of the big four to confirm that massive AI spending was very much on the cards for 2025.

Here’s the scorecard for 2025 Capex, totaling over $320Bn. A few months back, estimates were swirling for $250 to $275. Goldman had circulated $300Bn in total Capex for the year, and these four have already planned more.

Amazon $105Bn
Microsoft $80Bn
Alphabet $75Bn
Meta $60 to $65Bn
Total $320Bn

The earnings call discussed DeepSeek R1 and the lower AI cost structure that it may presage, with the possibility of lower revenue for AI cloud services.

“We have never seen that to be the case,” Amazon CEO Andy Jassy said on the call. “What happens is companies will spend a lot less per unit of infrastructure, and that is very, very useful for their businesses. But then they get excited about what else they could build that they always thought was cost prohibitive before, and they usually end up spending a lot more in total on technology once you make the per unit cost less.”
Amazon plans to spend heavily on custom silicon and focus on inference as well besides buying Blackwells by the truckload.

Q3-FY2025

Marvell reported impressive Q3 results that beat revenue estimates by 4% and adjusted EPS estimates by 5.5%, led by strong AI demand. FQ3 revenue accelerated to 6.9% YoY and 19.1% QoQ growth to $1.52 billion, helped by a stronger-than-expected ramp of the AI custom silicon business.

For the next quarter, management expects revenue to grow to 26.2% YoY and 18.7% QoQ to $1.8 billion at the midpoint. The Q4 guide beats revenue estimates by 9.1% and adjusted EPS estimates by 13.5%. Management expects to significantly exceed the full-year AI revenue target of $1.5 billion and indicated that it could easily beat the FY2026 AI revenue target of $2.5 billion.

Marvell has other segments, which account for 27% of the business that are not performing as well, but they’re going full steam ahead to focus on the custom silicon business and expect total data center to exceed 73% of revenue in the future.

  • Adjusted operating margin – 29.7% V 29.8% last year, and better than the management guide of 28.9%.
  • Management guidance for Q4 is even higher at 33%.
  • Adjusted net income – $373 Mn or 24.6% of revenue compared to $354.1 Mn or 25% of revenue last year.
  • Management has also committed to GAAP profitability in Q4, and continued improvements.

Custom Silicon – There are estimates of a TAM (Total Addressable Market) of $42 billion for custom silicon by CY2028, of which Marvell could take 20% market share or $8Bn of the custom silicon AI opportunity, I suspect we will see a new forecast when the company can more openly talk about an official announcement. On the networking side, the TAM is another $31 billion.

“Oppenheimer analyst Rick Schafer thinks that each of Marvell’s four custom chips could achieve $1 billion in sales next year. Production is already ramping up on the Trainium chip for Amazon, along with the Axion chip for the Google unit of Alphabet. Another Amazon chip, the Inferentia, should start production in 2025. Toward the end of next year, deliveries will begin on Microsoft’s Maia-2, which Schafer hopes will achieve the largest sales of all.”

Key weaknesses and challenges

Marvell carries $4Bn in legacy debt, which will weigh on its valuation.

The stock is already up 70% in the past year, and is volatile – it dropped $26 from $126 after the DeepSeek and tariffs scare.

Custom silicon, ASICs (Application Specific Integrated Circuits) have strong competition from the likes of Broadcom and everyone is chasing market leader Nvidia. Custom silicon as the name suggests is not widely used like an Nvidia GPU and will encounter more difficult sales cycles and buying programs.

Drops in AI buying from data center giants will hurt Marvell.

Over 50% of Marvell’s revenue comes from China, and it could become a victim of a trade war.

Valuation: The stock is selling for a P/E of 43, with earnings growth of 80% in FY2025 and 30% after that for the next two years – that is reasonable. It has a P/S ratio of 12.6, with growth of 25%. It’s a bit expensive on the sales metric, but with AI taking an even larger share of the revenue pie, this multiple could increase.

Categories
AI Cloud Service Providers Industry Semiconductors Stocks

Hyperscalers, Meta and Microsoft Confirm Massive Capex Plans

Blackwell

Meta (META) has committed to $60-65Bn of Capex and Microsoft (MSFT) $80Bn: After the DeepSeek revelations, this is a great sign of confidence for Nvidia (NVDA), Broadcom (AVGO), and Marvel (MRVL) and other semiconductor companies. Nvidia, Broadcom, and Marvell should continue to see solid demand in 2025.

Meta CEO, Mark Zuckerberg also mentioned that one of the advantages that Meta has (and other US firms by that same rationale) is that they will have a continuous supply of chips, which DeepSeek will not have, and the likes of US customers like Meta will easily outperform when it comes to scaling and servicing customers. (They will fine-tune Capex between training and inference). Meta would be looking at custom silicon as well for other workloads, which will help Broadcom and Marvell.

Meta executives specifically called out a machine-learning system designed jointly with Nvidia as one of the factors driving better-personalized advertising. This is a good partnership and I don’t see it getting derailed anytime soon.

Meta also talked about how squarely focused they were on software and algorithm improvements. Better inference models are the natural progression and the end goal of AI. The goal is to make AI pervasive in all kinds of apps for consumers/businesses/medical breakthroughs, and so on. For that to happen you still need scalable computing power to reach a threshold when the models have been trained enough to provide better inference and/or be generative enough, to do it for a specific domain or area of expertise.

This is the tip of the iceberg, we’re not anywhere close to reducing the spend. Most forecasts that I looked at saw data center training spend growth slowing down only in 2026, and then spending on inference growing at a slower speed. Nvidia’s consensus revenue forecasts show a 50% revenue gain in 2025 and 25% thereafter, so we still have a long way to go.

I also read that Nvidia’s GPUs are doing 40% of inference work, they’re very much on the ball on inference.

The DeepSeek impact: If DeepSeek’s breakthrough in smarter inference were announced by a non-Chinese or an American company and if they hadn’t claimed a cheaper cost, it wouldn’t have made the impact it did.  The surprise element was the reported total spend, and the claim that they didn’t have access to GPUs – it was meant to shock and awe and create cracks in the massive spending ecosystem, which it is doing. But the reported total spend or not using high GPUs doesn’t seem plausible, at least to me. Here’s my earlier article detailing some of the reasons. The Chinese government subsidized every export entry to the world, from furniture to electric vehicles, so why not this one? That has been their regular go-to-market strategy. 

Cheaper LLMs are not a plug-and-play replacement. They will still require significant investment and expertise to train and create an effective inference model. I think the GPU requirements will not diminish because you need GPUs for training and time scaling, smarter software will still need to distill data. 

Just as a number aiming at a 10x reduction in cost is a good target but it will compromise quality and performance. Eventually, the lower-tier market will get crowded and commoditized – democratized if you will, which may require cheaper versions of hardware and architecture from AI chip designers, as an opportunity to serve lower-tier customers.

American companies will have to work harder, for sure – customers want cheap (Databricks’ CEO’s phone hasn’t stopped ringing for alternative solutions) unless they TikTok this one as well…..

Categories
AI Cloud Service Providers Industry Semiconductors Stocks

DeepSeek Hasn’t Deep-Sixed Nvidia

01/28/2025

Here is my understanding of the DeepSeek breakthrough and its repercussions on the AI ecosystem

DeepSeek used “Time scaling” effectively, which allows their r1 model to think deeper at the inference phase. By using more power instead of coming up with the answer immediately, the model will take longer to research for a better solution and then answer the query, better than existing models.

How did the model get to that level of efficiency?

DeepSeek used a lot of interesting and effective techniques to make better use of its resources, and this article from NextPlatform does an excellent job with the details.

Besides effective time scaling the model distilled the answers from other models including ChatGPT’s models.

What does that mean for the future of AGI, AI, ASI, and so on?

Time scaling will be adopted more frequently, and tech leaders across Silicon Valley are responding to improve their methods as cost-effectively as possible. That is the logical and sequential next step – for AI to be any good, it was always superior inference that was going to be the differentiator and value addition.

Time scaling can be done at the edge as the software gets smarter.

If the software gets smarter, will it require more GPUs?

I think the GPU requirements will not diminish because you need GPUs for training and time scaling, smarter software will still need to distill data. 

Cheaper LLMs are not a plug-and-play replacement. They will still require significant investment and expertise to train and create an effective inference model. Just as a number aiming at a 10x reduction in cost is a good target but it will compromise quality and performance. Eventually, the lower-tier market will get crowded and commoditized – democratized if you will, which may require cheaper versions of hardware and architecture from AI chip designers, as an opportunity to serve lower-tier customers.

Inferencing

Over time, yes inference will become more important – Nvidia has been talking about the scaling law, which diminishes the role of training and the need to get smarter inference for a long time. They are working on this as well, I even suspect that the $3,000 Digits they showcased for edge computing will provide some of the power needed.

Reducing variable costs per token/query is huge: The variable cost will reduce, which is a huge boon to the AI industry, previously retrieving and answering tokens cost more than the entire monthly subscription to ChatGPT or Gemini.

From Gavin Baker on X on APIs and Costs:

R1 from DeepSeek seems to have done that, “r1 is cheaper and more efficient to inference than o1 (ChatGPT). r1 costs 93% less to *use* than o1 per each API, can be run locally on a high end work station and does not seem to have hit any rate limits which is wild.

However, “Batching massively lowers costs and more compute increases tokens/second so still advantages to inference in the cloud.”

It is comparable to o1 from a quality perspective although lags o3.

There were real algorithmic breakthroughs that led to it being dramatically more efficient both to train and inference.  

On training costs and real costs:

Training in FP8, MLA and multi-token prediction are significant.  It is easy to verify that the r1 training run only cost $6m.

The general consensus is that the “REAL” costs with the DeepSeek model much larger than the $6Mn given for the r1 training run.

Omitted are:

Hundreds of millions of dollars on prior research and has access to much larger clusters.

Deepseek likely had more than 2048 H800s;  An equivalently smart team can’t just spin up a 2000 GPU cluster and train r1 from scratch with $6m.  

There was a lot of distillation – i.e. it is unlikely they could have trained this without unhindered access to GPT-4o and o1, which is ironical because you’re banning the GPU’s but giving access to distill leading edge American models….Why buy the cow when you can get the milk for free?

The NextPlatform too expressed doubts about DeepSeek’s resources

We are very skeptical that the V3 model was trained from scratch on such a small cluster.

A schedule of geographical revenues for Nvidia’s Q3-FY2025 showed 15% of Nvidia’s or over $4Bn revenue “sold” to Singapore, with the caveat that it may not be the ultimate destination, which also creates doubts that DeepSeek may have gotten access to Nvidia’s higher-end GPUs despite the US export ban or stockpiled them before the ban. 

Better software and inference is the way of the future

As one of the AI vendors at CES told me, she had the algorithms to answer customer questions and provide analytical insides at the edge for several customers – they have the data from their customers and the software, but they couldn’t scale because AWS was charging them too much for cloud GPU usage when they didn’t need that much power. So besides r1’s breakthrough in AGI, this movement has been afoot for a while, and this will spur investment and innovation in inference. We will definitely continue to see demand for high-end Blackwell GPUs to train data and create better models for at least the next 18 months to 24 months after which the focus should shift to inference and as Nvidia’s CEO said, 40% of their GPUs are already being used for inference.