WE ARE BUILDING AI LIKE IT IS 1969

The efficient alternative was always there. The race just drowned it out.

The Saturn V was extraordinary engineering. It was also entirely disposable. You lit the fuse, it burned $185 million in today's money in a single flight, most of it fell into the ocean, and you built another one. The mission was achieved. The model was insane.

The alternative thinking existed from the beginning. The Space Shuttle was the first serious attempt at reusability—flawed, still expensive, but the architecture was right. Mechanical launchers for satellites. Rail-based systems. Nuclear propulsion concepts. Aerospace engineers were imagining efficient alternatives to the disposable rocket since the 1970s.

The problem was never imagination. It was that the race logic kept funding the Saturn V model because that's what won prestige, contracts, and national headlines. The efficient alternative didn't need inventing. It needed the race to exhaust itself first.

We are in the Saturn V era of AI. The efficient alternative already exists. The race is starting to run out of runway. And the rocket being built isn't as permanent as the headlines suggest.

𝗧𝗵𝗲 𝗔𝗜 𝘀𝗽𝗮𝗰𝗲 𝗿𝗮𝗰𝗲 𝗻𝗼𝗯𝗼𝗱𝘆 𝘃𝗼𝘁𝗲𝗱 𝗳𝗼𝗿

Between $3.7 and $7.9 trillion in data centre investment is projected by 2030—just to build the infrastructure, before a single query is run. By the end of this decade those facilities will consume power equivalent to 180 million US homes. Large data centres use up to 5 million gallons of water per day for cooling. Energy bills in communities hosting these facilities are already rising.

The four largest tech companies spent $413 billion on AI infrastructure in 2025 alone. Guidance suggests $600-700 billion in 2026. The construction is accelerating regardless of whether local communities want it, can power it, or can afford the grid consequences.

Nobody voted for this. It was decided between tech giants and governments competing for AI supremacy in boardrooms and policy corridors most people will never enter. The public inherited the consequences, as they usually do.

This is the Saturn V model applied to intelligence. More compute. Bigger clusters. Faster chips. Win by scale. Justify the cost with the race. Build more infrastructure to meet demand. Repeat.

But call it what it actually is before we go further.

𝗧𝗵𝗲 𝗺𝘆𝘁𝗵 𝗼𝗳 𝗔𝗜 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲

When politicians and tech executives say infrastructure, most people picture a power station. A railway network. A water system. Things built to last decades, amortised across generations of use, serving multiple purposes long after the original investment is made. Things that justify their cost over time.

AI data centre infrastructure is not that.

The hardware inside—the GPUs (specialised chips that power AI processing), the servers, the networking equipment—has a functional lifespan of three to five years before it is commercially obsolete. The chips being installed today will be e-waste before this decade ends. The buildings may stand longer, but the technology stack they house is on a permanent replacement cycle that never stops consuming capital.

You are not building a railway. You are building a Saturn V factory. One that has to keep manufacturing rockets it burns in a single flight—except now the rockets are also obsolete before they land.

The $7 trillion headline figure implies permanence. The reality is churn. You do not build $7 trillion of AI infrastructure once. You build it, replace it, build it again, replace it again, at accelerating cost, just to stay current with the model generation being released that year. The depreciation curve is brutal and almost never appears in the investment projections being quoted to justify the race.

When the next generation of chips arrives—and it arrives on a roughly eighteen-month cycle—the current installation is not legacy infrastructure serving reduced capacity. It is a stranded asset competing against hardware that is faster, more efficient, and cheaper to run. The race does not pause for the return on the last investment to materialise.

This is the Saturn V model in full. Disposable rocket. Disposable infrastructure. Enormous capital consumed in service of a race whose finish line keeps moving.

And some of the most significant players in the race have noticed.

Apple and Microsoft—two of the largest technology companies on earth—are quietly repositioning. Rather than doubling down on owning the model layer, both are moving toward becoming the pipe: the operating system, the device, the productivity platform through which any AI model can flow. Microsoft's M365 now connects to third party models including Claude. Apple's Siri is opening to external AI integration. Neither is exiting the race. Both are declining to be defined by it. They are choosing the railway over the rocket—infrastructure that survives whoever wins the model war, that earns margin regardless of which AI sits inside it.

That is not retreat. That is the hedge of players who have watched enough technology cycles to know that platform ownership outlasts product cycles. And it is a quiet acknowledgement that the Saturn V model is not where experienced capital wants to be locked long term.

𝗦𝗼𝗺𝗲𝗼𝗻𝗲 𝗷𝘂𝘀𝘁 𝘀𝘁𝗮𝗿𝘁𝗲𝗱 𝗳𝘂𝗻𝗱𝗶𝗻𝗴 𝘁𝗵𝗲 𝗮𝗹𝘁𝗲𝗿𝗻𝗮𝘁𝗶𝘃𝗲

Last week Google released Gemma 4.

Gemma 4 is an open weights model—meaning the architecture and parameters are publicly released, free for anyone to use, modify, and run without paying Google anything. Critically it runs on-device. On a phone. Offline. Without routing your data through a data centre on another continent. Without a subscription. Without contributing to the $7 trillion infrastructure bet. Without adding to the energy and water consumption of a community that had no say in hosting the facility.

On performance benchmarks—standardised tests measuring reasoning, language understanding, and problem-solving—Gemma 4 competes with models many times its size requiring vast server infrastructure to run. It is not identical to frontier models like Claude Opus or GPT-4o. For deep research, extended reasoning, and complex enterprise work those still hold advantages. But for an enormous range of professional tasks it is good enough. And the marginal cost of running it is zero. It does not depreciate. It does not require a replacement cycle funded by someone else's capital raise.

The irony worth naming directly: Gemma 4 was released by Google—one of the largest players in the data centre space race, spending tens of billions annually on the Saturn V model. The company most invested in the infrastructure arms race simultaneously released the architecture that begins to undermine it.

That is not confusion. That is the hedge of a player smart enough to own multiple positions simultaneously. Google is funding the giant disposable rocket and the efficient alternative at the same time. It intends to win whichever model the market eventually selects.

𝗧𝗵𝗲 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗺𝗼𝗱𝗲𝗹 𝗶𝗻 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗲

For creatives and small businesses the implications are significant and largely undiscussed.

If capable AI inference—the process of running a model to generate a response—approaches zero cost at local compute level, you are no longer dependent on a subscription currently priced below its true cost and due for correction. You are no longer routing sensitive client work through third party servers you do not control. You are no longer one platform shutdown, acquisition, or pricing correction away from losing a tool embedded in your workflow. And you are no longer an invisible participant in an infrastructure bet you never agreed to fund through your energy bill.

The creative community has been quietly building on this logic already. ComfyUI—an open source tool widely used for generative image and video work—operates entirely locally. No subscription, no data leaving your machine, no per-generation cost once installed. Blendworks is exploring similar ground at the professional end, bridging broadcast-quality motion and generative AI in ways that don't require routing everything through a hyperscaler. While mainstream conversation fixated on Midjourney and Sora, a significant creative community built serious infrastructure on the efficient model. The technology that changes daily creative practice rarely arrives with the loudest launch.

Gemma 4 extends that logic to language. The assistant, the analyst, the drafting partner—on your device, privately, at no ongoing cost, on hardware you already own and were already planning to replace on your own schedule.

The frontier model labs will not be killed by this. But they will have to justify what they charge for in a world where good enough is free, private, and requires no infrastructure that anybody had to vote for. That is a harder conversation than the one the data centre narrative has been having.

𝗧𝘄𝗼 𝗔𝗜𝘀 𝗮𝗿𝗲 𝗲𝗺𝗲𝗿𝗴𝗶𝗻𝗴

The Saturn V AI—powerful, enormously expensive to run honestly, built on short-lived hardware dressed as permanent infrastructure, dependent on decisions made without public consent, heading toward a pricing correction the industry is not advertising.

And the efficient AI—smaller, local, private, free to run, built on hardware people already own, improving faster than the brute force narrative wants to acknowledge, and built on thinking that was never absent. Just underfunded.

The space race produced the moon landing. The technology that changed daily life came later, from a different direction, built by people who had been asking the same questions for decades and finally had permission to answer them properly.

The efficient alternative was always there.

The race just drowned it out. Until now.

Sources: JLL 2026 Global Data Centre Outlook, McKinsey Global Institute data centre investment projections, Brookings Institution data centre energy and water analysis, Google Gemma 4 release documentation