Technology Validation
In another surprise, AMD announced a major AI chip deal with Meta Platforms (META). The plan is for the social media giant to deploy up to 6 GW of AMD Instinct chips to power AI infrastructure.
Meta plans to deploy the MI450 GPU in the 2H’26 to support the first GW deployment. The deal also includes the use of EYPC CPUs running ROCm software built with the Helios rack-scale architecture.
The major AI chip deal is structured similar to the one with OpenAI, where AMD will issue 160 million warrants. The warrants have performance requirements to be exercised by Meta Platforms, including the purchase of GPUs and the AMD stock reaching $600.
AMD estimates the deal brings in over $100 billion in total revenues. As the WSJ explains, the MI450 is easier to customize, giving Meta more control to target these GPUs toward inference functions key to the social media business.
Ultimately, these massive deals help to validate the AI GPU technology. AMD will sign up even more customers with the technology validated by OpenAI and Meta.
Path to $600
AMD now has 2 massive contracts with major hyperscalers with combined values topping $200 billion. While investors are focused on AMD handing out warrants like candy, most people don’t focus enough on how AMD has to reach $600 in order for OpenAI and now Meta Platforms to cash in all of the warrants.
On the partnership call, CFO Jean Hu reinforced how this deal pushes the chip company closer to the $20+ EPS target:
This partnership marks another significant step forward in delivering our ambitious long-term financial model, including greater than 80% CAGR of our data center AI business and generating more than $20 in annual earnings per share within the next 3 to 5 years.
While investors question the company giving up 10% of the outstanding stock in both deals, AMD is ensuring the company is solidly included in the massive AI future. OpenAI and Meta will both be entrenched in the future development plans, leading to far more revenues in follow-on contracts while financially tied to the success of AMD via a potential ownership of 160 million shares. At $600, the ownership positions could reach $96 million.
What’s a “warrant”? …minimum twenty pointless characters needed to post…
Might actually be $96 Billion.
A warrant is similar to an option, IIRC.
Mike
Gemini sayeth:
Stock warrants are long-term contracts issued directly by a company that give investors the right—but not the obligation—to buy company stock at a set price (strike price) before a specific expiration date. Used to raise capital or “sweeten” financing deals, warrants can dilute existing shares when exercised.
The Long-Term Stock Exchange +4
Key Aspects of Stock Warrants:
-
Structure: They function similarly to call options (allowing purchases) or put warrants (allowing sales), but are issued by the company itself rather than traded on an exchange.
-
Expiration: Warrant lifespans are generally long, often ranging from two to 10 years.
-
Exercise Price: This fixed price is typically higher than the market price at the time of issuance.
-
Dilution: When exercised, the company issues new shares, which can dilute the value of existing shares.
-
No Voting/Dividends: Warrant holders do not have voting rights or receive dividends until they exercise their warrants to own the underlying stock.
Investopedia +6
Warrants vs. Options
-
Issuance: Warrants are issued by the company (often to investors or lenders as a “sweetener”), whereas options are created between investors on an exchange.
-
Expiration: Warrants usually have much longer terms (years) compared to options (months).
-
Impact: Exercising warrants increases the company’s outstanding share count, while exercising options does not.
Investopedia +4
Why Companies Issue Warrants
Companies often issue warrants to attract investors, particularly when offering corporate bonds or preferred stock, as they provide potential upside if the stock price increases. They are also used to make financing deals more attractive.
SmartAsset +1
Thanks. Yeah, I should have run it through AI. Gemini happens to be my go-to. Seems pretty decent though I haven’t truly compared it with others like Grok, etc. I use it’s “Thinking” mode, which is one step above “Fast” but is still free.
–
Aside: Just yesterday I asked Gemini “Does the infinite power tower of the successively smaller fractions 1/2, 1/3, 1/4, 1/5, etc. converge?” In under a minute, it gave me an excellent and correct answer:
Unlike a power tower of a constant fraction like $1/2$, the tower of successively smaller fractions ($1/2, 1/3, 1/4, \dots$) does not converge to a single value. Instead, it diverges by oscillation.
As you add more layers to the tower, the result bounces back and forth between two distinct numbers.
The Two Limits
If you calculate the tower for a large number of terms ($n$), the result will settle into a “loop” depending on whether you have an odd or even number of fractions:
Even-numbered towers (e.g., $n = 100, 200, \dots$) approach $\approx 0.690$.
Odd-numbered towers (e.g., $n = 101, 201, \dots$) approach $\approx 0.658$.
Because the sequence never settles on one final number, mathematicians say the limit does not exist.
Why does this happen?
In our previous discussion, we noted that for a constant tower $x^{x^{x^{\dots}}}$ to converge, the base $x$ must be greater than $e^{-e} \approx 0.066$.
Falling below the limit: In your sequence ($1/2, 1/3, 1/4, \dots$), the fractions eventually become smaller than $0.066$. For example, $1/16 \approx 0.0625$.
Loss of Stability: Once the “top” of your tower includes these tiny fractions, the entire structure becomes unstable.
The “Yo-Yo” Effect: A very small number at the top of a power tower acts like an exponent that pulls the value toward $1$. But then that value (near $1$) acts as an exponent for the number below it, pulling it back down toward its base. This creates a permanent “tug-of-war” that results in the two values you see above.
Comparison: Constant vs. Successive
Tower Type Base(s) Result Constant $1/2, 1/2, 1/2, \dots$ Converges to $\approx 0.641$ Successive $1/2, 1/3, 1/4, \dots$ Oscillates between $\approx 0.658$ and $\approx 0.690$
Interestingly, if the fractions decreased extremely fast (much faster than $1/n$), the tower could potentially converge again, but the natural harmonic sequence $1/n$ is just “slow” enough to keep the oscillation alive forever.
My Perl one-liner confirms the two value oscillation:
C:\Users\jla81>perl -e "$n=200; $r=1/$n; $r=(1/$n)**$r while --$n>1; print $r"
0.658365599266331
C:\Users\jla81>perl -e "$n=201; $r=1/$n; $r=(1/$n)**$r while --$n>1; print $r"
0.690347126114964
C:\Users\jla81>perl -e "$n=202; $r=1/$n; $r=(1/$n)**$r while --$n>1; print $r"
0.658365599266331
C:\Users\jla81>perl -e "$n=203; $r=1/$n; $r=(1/$n)**$r while --$n>1; print $r"
0.690347126114964