Tokenomics - the Elon Musk worldview as rocket man watches AI costs soar into orbit, but ...

Eyes skyward! No-one is immune from the reach of tokenomics; it’s just that some don’t care about it as much as others! Take Elon Musk’s SpaceX, Last month, the firm released version 4.5 of its Grok gen AI bot which had a major impact - token consumption tripled in its wake! But if you’ve pulled off a ludicrously successful IPO that so bewitched Wall Street that investors cheerfully looked beyond basic fiscal realities, turning the founder into a trillionaire in the process, then a mere tripling of token consumption isn’t something that’s about to put you off your stride. Certainly SpaceX has no intention of slowing down its spending on AI, quite the opposite in fact. For its latest fiscal quarter, CapEx jumped sixfold to $18.4 billion, of which $15.8 billion went on AI compute resources, up from $7.7 billion in Q1, even as its AI business turned in an operating loss of $1.3 billion. SpaceX CFO Bret Johnsen confirms that CapEx will remain at similar levels for the next couple of quarters as the company invests in improving AI models and expanding its AI infrastructure, insisting that despite the eye-watering sums involved, the company is being “efficient” in its spending: On the AI compute side, we’re able to deploy capital in such a way that we’re getting less than a one-year payback. Spend, spend, spend This year alone, SpaceX has signed compute capacity agreements with Google and Anthropic worth $920 million and $1.25 billion per month, respectively. But the need for more and more and more compute remains unsatiated. For Musk himself, the prize is definitely worth the outlay, whatever the size: It's kind of hard for people to wrap their minds around this, but it's not out of the question that at some point, Starlink will deliver a majority of the world's internet, at least in countries where we're allowed to operate, which is the vast majority of countries. This is important to bear in mind, and it's not in the infinite future; it’s less than 10 years. With the advent of AI and humanoid robotics and vehicle robotics and just a massive number of robots, the appetite for bandwidth will be much greater than it has been in the past. Think of a human consuming or producing perhaps a few hundred bits-per-second of output. Actually, the average human is outputting less than one bit-per-second over the course of a day, which is 86,400 seconds. Computers easily do billions of bits per second continuously over a 24-hour period. This is why I would expect the appetite for bandwidth to grow dramatically with the growth of AI and robotics. And, really, I think Starlink is the only thing that can actually service that bandwidth. Regarding AI, Grok 4.5 was “a huge improvement”, but there is more to be done, he goes on: We have Grok 4.6 coming out probably next week, and then Grok 4.7 is about three or four weeks from today, and we expect the cadence of AI development to improve dramatically. And then with Grok 5, which should be out before the end of this year, we will be incorporating the entire corpus of SpaceX data. Basically all the data that SpaceX has ever produced, which is a tremendous amount over the course of a quarter century, will be incorporated into Grok training. Compute demand So action needs to be taken to provide the compute necessary to realize all this, he argues: We are providing compute to others, and we are building and deploying compute, I think, faster, or our rate of growth certainly is faster, than anyone else. And our efficiency of compute deployment, I think, is also the highest. We expect to end this year with over two gigawatts of compute. Probably our cumulative compute online by the end of next year will be several times higher, maybe closer to 10 gigawatts of compute than five gigawatts of compute. Having said that, he revises his own estimate as actually too conservative: We're actually aiming to far exceed that gigawatt number in terms of power online, power cooling and electrical equipment. Our tentative target is to actually have 20 gigawatts at the power and cooling level online by the end of next year. Now I don't think we're going to achieve 20 gigawatts, but we want to have a series of projects that cumulatively come to 20 gigawatts by the end of next year. Some of them won't pan out exactly on time, but I would expect that we still probably have at the power plant level, something close to 15 gigawatts. So, assuming that maybe a quarter of the projects take longer than expected, our goal is to have far more power, cooling and electrical equipment than we have GPUs. That's a logical thing to do given the relative expense of GPUs versus balance of system. On the subject of GPUs, Musk is keen to highlight an important strategic decision: Going forward, we've decided to build exclusively on NVIDIA because we think the Vera Rubin architecture is the best architecture. We think it's the best AI computer, and we greatly value our close cooperation and partnership on many levels with NVIDIA. So we're exclusive to NVIDIA. This new capacity coming online will obviously further benefit the development of Grok..Our understanding with NVIDIA is that we will receive a very significant percent of their GPUs next year. SpaceX as a company is “incredibly good” at hardware, he boasts, and it isn’t, as the saying goes, rocket science to see that in action: Rocket science is literally our daily business. Rocket science is an idiomatic expression for extreme technological difficulty, and there's a reason for that because let me tell you what rockets desperately want to do every flight - they desperately want to blow themselves into tiny pieces. The engineering struggle is to convince the rocket not to blow itself into tiny pieces and actually deliver payload to orbit. If you apply a small amount of the engineering skill that goes into making gigantic rockets reach orbit on a regular basis reliably to terrestrial data centers, this is like, frankly, the New York Yankees going in and playing a little league team. It's kind of ridiculous, frankly. Terrestrial data centers are a trivial problem compared to making gigantic reusable rockets and which are launched frequently. So to that end: We're taking a small amount of the expertise that we use for rockets and satellites and applying that to scaling terrestrial data centers. And we're finding that even a small amount of what we've learned, building rockets, which are incredibly difficult, applied to data centers yields tremendous benefits. So yes, we expect to have far in excess of the power cooling that's needed. Splashdown All of this has economic implication, he posits: What we're seeing with AI is that intelligence-per-watt is increasing rapidly. So that means the usefulness-per-watt is increasing rapidly. Anyone who's a close follower of the AI arena, you start seeing smaller and smaller models do more and more amazing things. There’s an argument that we may get to the point where even a [NVIDIA] H100 GPU is able to deliver useful AI that's comparable to a human engineer. Based on current trends, I think that's not an impossible outcome. So then it's like how useful would that be? I think the usefulness that is going to be extremely high. And the ongoing speed of evolution across the AI sector means it’s critical to remain flexible, he says: If you think of where was AI 12 months ago? Twelve months ago have been August last year, before the advent of Claude 4.5, which should be considered one of the milestones and the credit [goes] to Anthropic. It was, I would say, a shock to the system how good Claude 4.5 was, and that was 10 months ago, basically. And if you say, what about an AI [from] two years ago? I think an AI two years ago feels like something that should be in a museum. It's a relic. So I think by the end of next year, it's not clear to me that there's anything that’s digital at least that AI won't be able to do, based on the current rate of improvement. The limiting factor here is the rate at which logic and memory can be produced, he argues, and that has bottom line budgetary implications for enterprises: Memory output is increasing by around 20% per year. Normally, that would be fantastically fast and amazing for any large mature industry. But ask yourself, is demand increasing by 20% a year? No, demand is increasing by 200% a year, maybe higher. So if you've got demand increasing much faster than supply, then Economics 101 would suggest that the price increases. It does not decrease. My take As Musk was making his latest comments, parts of a SpaceX rocket crashed into the Moon It’s tempting to view this collision as a metaphor for the the company’s plunging share price as investors stirred from their blissed out post-IPO state, awoken perhaps by those CapEx estimates which were enough to scare even the most long-termist of horses, let alone the short-termists on Wall Street looking for their next reason to panic. But Musk will hold the line, no doubt about that. He may have lost his paper trillionaire status, for now, but he doesn’t have to care about rising bills when there’s a mission of literally galactic proportions to be wrestled with. That’s fine and dandy and capitalism in action, of course, but it’s not something that’s going to provide a great deal of comfort for the enterprise exec reeling from the size of their latest AI bill and looking in vain for all that value creation that the out-of-control hype cycle promised would be easy pickings. Eyes up to the heavens, everyone - the future is out there...if you can afford it.