Some thoughts on AI - Oracle and others

I have held Oracle for well over 10 years. I finally sold last week.

Key reason was, in the world of AI, Oracle is becoming a different company.

Previously, it was a classic enterprise software business. Capital light. Grew steadily year after year. Ellison kept increasing his stake in the company through buybacks. In 2023, Oracle had sales of $49bn and had net profit of $14bn. It spent about $9bn in capX in 2023.

Then the era of AI..... Oracle/Larry decided to go for it. It is looking to scale up its cloud/OCI division through its partnership with OpenAI. In 2025, its capX was $56bn. In 2026/2027, it is spending $95bn and $116bn. As we can see, its core enterprise software business would not be capable of funding its cloud/AI investment. It is borrowing heavily to fund its capX program.

Normally, I am OK with investing for the long haul. But in this case, what is the value add of being a cloud partner to OpenAI ? Who is capturing the value at the end of the day ? Put it another way, the nature of business of Oracle has changed substantially in the era of AI.

Then, Oracle declared force majeure on its Jupiter project in New Mexico (2.45GW). Key reason was it was struggling to get permits for its gas pipeline to supply its Bloom Energy's fuel cells. Interesting that public sentiment has turned against AI. NIMBY is alive and well in USA.

Apart from challenges with the local community, the bond market is waking up. Bonds from Oracle is now yielding 8% or so.

By the way, 2.45GW in power requires almost $100bn in chips from Nvidia.

Staying in USA, we have SB Energy have to delay its IPO. It was originally targeting valuation of $50bn. See Pic 2 for what it has... It has 800MW that are under construction and will be online in 2027 (50MW) and 2028 (750MW). It also has contracted power of 8GW (not under construction) that is only expected to come online in 2031/2032. In 1H 26, it had sales of $139M. It also had net debts of $1.8bn.

Put it politely, SB Energy has nothing as of today. To build out its first 800MW, it needs around $32bn for the GPUs. Clearly, it needs to raise money (and a lot of it) to fund even the first 800MW. BUT my beef is does a few contracts of power supply + land worth $50bn when we are looking at capX of $30bn + $320bn to come. Admittedly, it is backed by backlog/offtake from OpenAI of $438bn.

And then the next question is... Can we really be comfortable that OpenAI is good for $438bn ?

By the way, we have our SB Energy in Australia... Firmus.....

I honestly don't know how to value these companies. I don't think anyone has a clue how much is OpenAI good for.. Despite Nvidia is happy to backstop (credit support) SB Energy GPUs, SB Energy (and Oracle) will be heavily reliant on the debt market.

Now.... For these companies to work... We need:

1. OpenAI to be a blockbuster success.. as in growing to few 100s of billions in the next few years.

2. Debt market to remain open and kind... BUT given where bond yield is trending... And data centre debts going >8% (latest Cleanspark DC deal backstop by Meta).....These are not cheap debts anymore and investors are thinking harder than before.

3. The force majeure at Oracle highlights that political/community is shifting.... If Oracle can't get the pipeline sorted, it can't power up its GPUs.

4. Then if we throw in construction delay (these are big projects and time line is few years' out) + amount of debts require, execution has to be perfect (or else)....

And then we have other risk factors to consider:

1. What is the useful life of Nvidia GPUs ? Can Vera Rubin become obsolete by any new chips/inference chips from elsewhere ?

2. OpenAI/Anthropic are "close" models..... Will open models dilute the TAM for OpenAI/Anthropic ? Afterall, not all the tasks require the latest and greatest.

3. What happens when AI moves to the end points (like PC) with Small Language Models (SLM). See Pic 3. Again, another potential force to steal the TAM from OpenAI/Anthropic.

If anyone says they know what is going to transpire, good luck to them :)

This is what we know... Oracle's Jupiter site has 2.45GW up in the air.. SB Energy has another 8GW.... All up, we are looking at potential sales of $400-500bn for Nvidia if these two companies can line up financing for their data centers. If not, Nvidia will need to find other customers to cover these shortfalls. Remember Wall Street is crazy about momentum... The moment the delta slows, the worry of AI peak will again keep investors wide awake.

BUT...

1. Channel checks from any semi companies in past 1-2 months => 2027 is looking good/secured.

2. The likes of Nebius and IREN are talking about higher GPU prices + healthy demand for GPUs next year.

Base case is 2027 remains a growth year for AI.... BUT there are undercurrents to be aware of....

A historical precedence to be aware of... During the internet bubble in 2000, when CLECs struggled with their IPOs, it was a sign of the top..... Anthropic is pushing out its IPO to November. OpenAI IPO is delayed to 2027...... Hmmmmm....... These companies need the IPO proceeds to pay for GPUs coming online next year.

Quickly.... Where does Bitdeer fit in ? Tydal will come online in Q4/Q1 27... Contracted for over 10 years with Volta. The key risk is construction delay here as the contract is signed already. Then its neoclouds will be live in 2027.....Key is how these deals are structured.... As neocloud, half of the GPU capX is paid upfront by the customers. Payback of GPUs in 2-3 years....In the case of Bitdeer, it has relatively small amount of power (about 100MW in Malaysia (Q3 27) + 43MW in Norway + 65MW in USA in Oct/Dec 27. My suspicion the customers in Asia may not be OpenAI/Anthropic.. Once these contracts are signed, these "other" customers should be good for it. And Norway may capture the upcoming inference needs in Europe (you don't want to be in training too much I reckon). If it can't find customers for its 65MW in USA in late 2027, it is not the end of the world. Just plug its sealminers back online.

Bottom line is once these neocloud contracts are signed with different counter parties, Bitdeer will be able to pay off these GPUs in quick time.... Then FCF would spike in Yr3-5.