DD: The Upcoming "Token Factory" Supercycle. ~700K in GOOG, MSFT LEAPs
Last time I posted about my 30 bagger GOOGL LEAP play everyone flamed me for not posting until it had already partially materialized, so here I am right before I make my next massive play. I made that play because of similar macro-level considerations (the fact that Google's best model was slightly worse than ChatGPT, but 10-20x cheaper), as well as the market mis-pricing of what businesses AI was going to disrupt (search/ads disruption being overblown).
**Context:**
Currently, the market has crowned the winners of AI as the companies who are providing the infrastructure to the hyperscalers who are spending money hand over foot: semiconductors, GPU/TPUs, energy companies, electrical utilities, memory. In every investment report I have read, they explicitly emphasize that the "AI trade" is investment in the companies that profit from the absurd spending from the AI hyperscalers, and many have explicitly called out that token providers are currently not able to extract margins.
There is also a very pervasive (and I think accurate) narrative depressing the valuation of model creators and AI scalers overall: the reality that no AI model has a real moat. So in the long term, model creators (google, openAI, anthropic) have no real ability to demand high margins or monetize purely off of the quality of their model as switching costs are very low and quality is close. This narrative is debatable, and very well could be correct, but I think is missing the real point:
The value in the AI trade is going to shift from going 100% to the infrastructure players to going 50-50 to infrastructure and to datacenter hyperscalers (the owners of the "intelligence factories")
|Company|Est. Blackwell B200-equivalents OWNED and installed (not rented). Source: Epoch AI + others|
|:-|:-|
|**Google (Alphabet)**|**\~0.41M–0.67M**|
|**Microsoft (Azure)**|**\~0.24M–0.35M**|
|Meta|\~0.26M high uncertainty|
|Amazon (AWS)|\~0.10M–0.16M|
|xAI|\~0.15M|
|CoreWeave|\~0.08M–0.10M|
|Oracle (OCI)|\~0.03M–0.05M|
**THE DD:**
**Assumptions** (feel free to disagree with these):
* The latest generation of AI models: Gemini 3 Flash, Opus 4.6, GPT 5.3 Codex are genuinely able to provide economic value to companies essentially on their own. (Ask your best software engineer friends), or look at the fact that 4% of github commits are done by Claude Code today, but Anthropic captures a negligible amount of the software engineering labor market value
* The general enterprise world, small businesses, entrepreneurs, etc. have not yet realized this value and (in my opinion) will very soon
* Althought the raw cost of intelligence is decreasing ($/token decreasing, and the raw utility of models is rising), the ability to serve this intelligence as a service at massive scale takes \~2 years to get online. So the datacenter/intelligence business will likely become cyclical, like the memory business is today.
**Thesis**:
When the utility of the current generation of AI models begins to be realized, the price that GPU / datacenter owners can charge is no longer tied to cost, it's tied to the economic value of the models it is able to run. The unit of value is no longer the GPU itself, it's the tokens / intelligence that the datacenter can produce, and as the value of that intelligence rises, datacenters capture that value as pure profit. If you agree with my assumptions, then you will also agree that tokens are about to enter a massive supercycle as enterprises begin to realize the economic potential of AI. The price anthropic charges for API access is going to spike as they (and every other AI provider) are unable to serve demand, but that margin is not going to go primarily to them, it'll go to Google, who they rent their TPUs from.
You can think of a datacenter as a "token factory", and when the demand for intelligence spikes the margins of these factories do too, until enough factories come online to meet demand. We can debate about when that demand will be met (if ever) but to me the next 1-2 years are undoubtedly going to see increased prices for compute. Personally, I believe that the cost of compute will always be very very high in the future as I think there is literally infinite demand for intelligence, but that's a different argument that I don't think needs to be had here.
**Positions**:
* I'm holding onto my GOOG LEAPs, and rolling some that expire soon (with LTCG) into new 500C 2027-28 expiration LEAPS. These have dropped by an absurd \~400k since the earnings meltdown, but I took \~150k in gains for tax reasons and am still up about 25x from a $33k basis. Holding onto most of them as I maintain long term conviction.
* 65 GOOGL 380C Jan 15 27 \~$139,000
* 30 GOOGL 300C Jan 15 27 \~$153,000
* 50 GOOGL 365C Dec 17 27 \~$246,000
* 8 GOOGL 300 Call Jun 18 26 \~$38,000 (trimmed for tax reasons end of 2025)
* 15 GOOG 300 March 20 26 \~$28,000 (will roll at end of Feb)
* Going to slowly accumulate MSFT LEAPS, targeting 700-800C and 2028 expiration. Bought 6k worth today, will continue to do so every few days until I have about 20k.
* NEW: 10 MSFT 800C Jan 21 28 \~$6650
* Considering opening a META position, but have not moved yet. As I close my GOOGL leaps for LTCG between now and May, I will likely open substantially larger META/MSFT positions alongside new GOOGL LEAPs. Will post those positions when I make them, but I don't like to enter big positions all at once.
Feel free to comment if you disagree with my assumptions, I hold them as true based on my conversions with many at AI labs, my personal use of the models at work to do jobs I have never previously been able to do, and the opinions of many analysts in the space, but am very open to hearing why you may disagree.
(Post was banned, I'm assuming because I didn't have specific positions, so added them above)