End of the beginning or the beginning of the end?
There was a moment, not very long ago, when the arrival of artificial intelligence felt genuinely egalitarian. A programmer in Pune, a gig worker in Lagos, a bootstrapped founder in Bengaluru, all of them could, for the first time in the history of computing, access tools that previously belonged only to well-resourced research institutions and large corporations. The large language model, the open-source framework, the affordable GPU on a monthly cloud subscription: together they seemed to promise a real rebalancing of technological power. The future, it appeared, would belong to the intellectually agile, not merely the capital-heavy.
That window is closing. And the events of the past fortnight have made the mechanism of closure impossible to ignore.
The Checkpoint That Changed Everything
On 25 June 2026, it emerged that the Trump administration had asked OpenAI to limit the release of its next frontier model, GPT-5.6, to a small set of government-approved partners before any wider rollout. According to reporting by Axios, the request came from the White House’s Office of the National Cyber Director and the Office of Science and Technology Policy, with Commerce Secretary Howard Lutnick personally involved in ensuring all relevant parts of the federal government had tested and cleared the model. OpenAI CEO Sam Altman confirmed the arrangement in a memo to staff, acknowledging that federal authorities would be “approving access customer by customer during this preview period.”
This is not, on its face, an unreasonable precaution. A model described by government sources as having “Mythos-like” capability, placing it in the same power bracket as the most advanced AI systems currently in existence, is not a chatbot that writes sales emails. The national security concerns that animate Washington’s caution are at least partially legitimate. Cybersecurity experts have warned that models of this calibre could meaningfully accelerate sophisticated attacks against interconnected legacy infrastructure in sectors such as banking and utilities.
But the implications reach well beyond the narrow question of security review. For the first time in the history of commercial artificial intelligence, the United States government has pre-emptively inserted itself between a company and its customers, not after deployment, not in response to demonstrated harm, but before the model has reached the public at all. The “voluntary” framework that President Trump’s earlier AI security executive order described has revealed itself, in practice, to be something considerably less optional.
Altman’s own formulation was telling. “We’ve made clear to the US government,” he wrote to employees, “that this is not our preferred long-term model.” The phrasing of a man who has acquiesced to something he cannot, at this moment, refuse.
The Precedent: Fable 5 and the Pattern It Set
The GPT-5.6 checkpoint did not arrive without warning. Two weeks earlier, on 12 June 2026, the US Department of Commerce invoked export controls against Anthropic’s Fable 5 and Mythos 5, barring access to any foreign national, including foreign nationals on Anthropic’s own staff. Unable to verify user nationality at scale, Anthropic disabled both models for everyone, worldwide, immediately. The stated trigger was a narrow jailbreak that Anthropic contested as non-universal and no more capable than what was already available in OpenAI’s GPT-5.5, a model that faced no equivalent restriction. The asymmetry was noticed, and never explained.
Anthropic is separately in litigation with the same administration over its refusal to allow its models to be used for domestic surveillance and autonomous weapons. The company Washington sanctioned is also the company Washington is suing. These are not the conditions of a stable, rules-based regulatory environment. They are the conditions of a political economy in which a company’s relationship with government matters as much as the quality of its technology.
That is the pattern the GPT-5.6 episode has now confirmed and extended. But the gatekeeping does not begin at the frontier model. It begins at the hardware shop on the corner, and on 25 June 2026, it arrived there with unmistakable clarity.
RAMageddon: When the Data Centre Came for Your Laptop
On the same day that news of the GPT-5.6 checkpoint broke, Apple raised prices across fourteen product categories: all Macs, every iPad, Apple TV, HomePod, HomePod mini, and Vision Pro. Increases ranged from $30 on a HomePod mini to $1,300 on a high-end Mac Studio. The MacBook Neo, Apple’s entry-level laptop, went from $599 to $699. The MacBook Air, the machine of choice for millions of students and small entrepreneurs worldwide, now starts at $1,299, up from $1,099, a $200 increase overnight. Analysts at Counterpoint Research estimate that higher component costs could add a further $150 to $200 per iPhone when Apple’s autumn launch arrives. In a detail that carried its own particular sting, Apple simultaneously raised prices in its Certified Refurbished store, the refuge of the budget-conscious buyer, the student, the small entrepreneur who had learned to wait a generation and buy smart. That route closed on the same day the new prices went live.
Apple’s statement was uncharacteristically blunt: “We have never seen a component price increase this much, this quickly. We have shielded our customers from these increases so far, but we have now reached a point where we need to begin raising prices.” Tim Cook, in an interview with the Wall Street Journal the week prior, was starker still: “I’ve never seen anything like it in any area in over 40 years.” The industry term that has attached itself to this moment is “RAMageddon”, and it is worth understanding precisely what that means, because it is the thread that connects the hyperscaler boardroom to the student’s desk.
Random access memory is a critical component in virtually every computing device made. The world’s three dominant memory manufacturers — Micron, Samsung, and SK Hynix — have pivoted an increasing share of their production capacity toward high-bandwidth memory for AI data centres, the specialised chips that Nvidia needs in enormous quantities to build the GPU clusters that power frontier AI models. According to research firm TrendForce, conventional DRAM contract prices jumped around 90 percent in early 2026, and looked set to climb a further 58 to 63 percent in the following quarter, with NAND flash storage rising 70 to 75 percent. One PC maker reported that memory went from about 15 percent of a laptop’s parts cost to roughly 35 percent in a single quarter. When a core component nearly triples in cost share, even Apple’s famous supply-chain muscle runs out of room.
Apple was not alone on 25 June. Microsoft announced Xbox price increases of $100 to $150 per console and stated that storage and memory prices have more than doubled, with a further doubling expected by autumn 2027. Nintendo raised the Switch 2 by $50. PC gaming company Valve launched its new Steam Machine at a higher price than originally intended, citing the RAM shortage directly. Dell, Lenovo, Acer, Asus, Samsung, and Sony have all raised prices or confirmed that further increases are unavoidable. Analysts at Gartner and IDC expect the broader fallout to include fewer devices sold and the slow disappearance of the cheapest laptops from the market entirely.
The GPU market had already been signalling this for months. Between November 2025 and February 2026, DDR5 RAM prices jumped approximately 40 percent. SSD prices rose more than 70 percent over the same period. Consumer GPU prices rose roughly 15 percent overall, but the RTX 5070 Ti became 25 percent more expensive, a price increase of approximately $190 on a single card, now sitting above manufacturer’s suggested retail price in every region tracked. Major PC vendors, Lenovo, Dell, HP, Acer, and ASUS, have warned of 15 to 20 percent price hikes for systems shipping in the second half of 2026.
The person who feels this is not the hyperscaler. Amazon can absorb a 50 percent increase in memory costs without a board meeting. The person who feels this is the design student in Chandigarh budgeting for a MacBook that now costs $200 more than it did last week, the gig worker in Hyderabad whose ageing laptop is now harder to replace, the small entrepreneur in Kochi who had planned to upgrade her workstation this quarter and is now postponing. For these users, the promise of AI as a democratising technology has collided, at the hardware counter, with the physical reality of who AI infrastructure actually serves, and who it prices out.
This is not a coincidence. It is a direct transmission mechanism. The AI infrastructure arms race consumes memory. Memory becomes scarce. Consumer devices become more expensive. The individual who might have participated in the AI economy as a creator, a developer, a small-scale entrepreneur finds the entry cost rising precisely as the frontier recedes behind government checkpoints and hyperscaler reservation systems. The democratisation narrative and the hardware reality are now pointing in opposite directions.
The Infrastructure Architecture of Exclusion
The RAM shortage at the retail level is, in one sense, merely the visible surface of a far larger structural shift. The hyperscalers, Amazon, Microsoft, Google, Meta, are expected to spend a combined $725 billion on AI infrastructure in 2026 alone. Amazon has committed $200 billion to data centre expansion this year. Google between $175 and $185 billion. Meta between $115 and $135 billion. Goldman Sachs projects cumulative AI capital expenditure of approximately $7.6 trillion between 2026 and 2031. For context: the entire global semiconductor industry generated roughly $627 billion in revenue in 2024. The infrastructure buildout now dwarfs the industry that makes the chips inside the infrastructure.
The critical detail embedded in these numbers is not their scale. It is their sequencing. Capacity is being purchased, reserved, and contractually locked before it is built. The hyperscalers are not responding to demand, they are pre-empting it, building moats of compute that smaller players cannot cross. The consequence for the small founder, the bootstrapped researcher, or the public-sector AI initiative is structural exclusion from the frontier. Training a competitive large language model from scratch requires thousands of GPUs. Renting the necessary compute on a monthly basis at current rates costs more than most seed-stage startups will raise in their first year. The venture capital ecosystem, which might once have bridged this gap, has followed the money upward — towards infrastructure plays, hyperscaler partnerships, and the handful of frontier labs already embedded in the Washington policy conversation. The founder with intellectual capital but without institutional backing is not disrupting an incumbent. She is applying for a lease on infrastructure she will never own.
When the Landlord Also Controls the Regulator
What the GPT-5.6 episode has confirmed is that the concentration of AI infrastructure and the concentration of regulatory relationship are not separate phenomena. They are mutually reinforcing.
The five companies that dominate AI compute are also the five companies with the deepest ties to the US national security establishment. They hold the contracts, the clearances, the lobbyists, and the revolving-door relationships that shape how AI regulation is written and enforced. When the government decides to review a frontier model before release, it is not doing so through a transparent statutory process with published criteria, defined timelines, and equal application across all providers. It is doing so through informal conversations, described variously as “requests,” “guidance,” and “direction”, that larger incumbents with established government relationships can navigate more smoothly than challengers can.
The five companies that dominate AI compute are also the five companies with the deepest ties to the US national security establishment. They hold the contracts, the clearances, the lobbyists, and the revolving-door relationships that shape how AI regulation is written and enforced. When the government decides to review a frontier model before release, it is not doing so through a transparent statutory process with published criteria, defined timelines, and equal application across all providers. It is doing so through informal conversations, described variously as “requests,” “guidance,” and “direction”, that larger incumbents with established government relationships can navigate more smoothly than challengers can.
The effect is a regulatory environment that functions as a barrier to entry in addition to a safety mechanism. This need not be the result of deliberate design. It is sufficient that it is the result.
Anthropic put the problem plainly: the government should have the ability to block unsafe deployments, but only as part of a statutory process that is transparent, fair, clear, and grounded in technical facts. The key word is statutory. An ad hoc, relationship-mediated approval system is not a safety framework. It is a discretionary power, and discretionary powers benefit those with access.
The Geopolitical Recoil
The market’s response to the Anthropic ban was instructive, and not in a way that Washington appears to have anticipated. Chinese lab Z.ai’s shares jumped over 30 percent. DeepSeek closed a funding round of approximately $7.4 billion, the largest in its history. Demand for Chinese AI models on third-party access platforms reportedly overtook demand for US models in the days following the Anthropic order.
The lesson is not that China’s AI capabilities are superior. The lesson is that nations and organisations that invested in sovereign compute and open-weight model development now have options. Those that did not are discovering, in real time, what dependency costs.
Critics noted that restricting Chinese-origin researchers from accessing top-tier models, including the many who work at Anthropic and OpenAI, would likely incentivise their return to China, directly undermining the administration’s stated objective of maintaining an AI lead. The policy may be accelerating the very outcome it was designed to prevent. This is what happens when security policy is made faster than it is thought through.
The Broken Promise of Democratisation
It is worth pausing to recall what was promised.
The early narrative of the AI era, the one that powered the valuations, the recruitment pitches, and the policy enthusiasm of 2020 through 2023, was explicitly democratising. Intelligence would be commoditised. The tools of the research lab would reach the classroom, the clinic, the small business, the individual creator. The gig worker would have a personal analyst. The farmer in Punjab would have an agronomist. The first-generation entrepreneur without institutional connections would compete on the quality of her ideas rather than the depth of her capital.
That narrative was not entirely false. For a period, the open-source movement, the API economy, and the declining cost of inference made genuine access possible for a broader constituency than had ever previously had access to frontier technology. The moment was real.
What has changed is not the technology. NVIDIA’s next-generation Vera Rubin GPUs, expected to ship in Q3 2026, promise to reduce cost per token by a factor of ten. These are meaningful efficiency gains and they deserve acknowledgement. But efficiency gains within a system do not alter the ownership structure of that system. Cheaper rent is not ownership. A model that costs less to run, accessed through an API controlled by a company that has pre-cleared its relationship with the US government, served from infrastructure that five corporations have reserved years in advance, this is not democratised intelligence. It is toll-road intelligence. The road may be faster. The toll is higher. And you will never own the road.
Meanwhile, the student who might have built on that toll road cannot afford the on-ramp. Her MacBook cost $200 more this week than it did last week. Her GPU upgrade has been postponed indefinitely. Her cloud compute bill has risen without announcement. The AI revolution is proceeding, but she is watching it from the pavement.
A Note on India
For India, the reckoning has a particular character. The country possesses abundant intellectual capital, in mathematics, engineering, linguistics, and domain expertise across agriculture, healthcare, law, and governance, but remains dangerously thin on sovereign AI infrastructure. India’s GPU capacity, its ability to train frontier-class models domestically, and its data centre density relative to its population and ambitions are each a fraction of what the present moment demands.
The hardware price spiral compounds this vulnerability at the base layer. As Apple MacBooks, consumer GPUs, and PC components become more expensive in dollar terms, and India imports virtually all of this hardware, the aspiring AI developer, the tier-two city startup, and the rural digital entrepreneur face a double bind: rising device costs at the entry point and rising compute costs at the frontier. The IndiaAI Mission and the proposed national compute infrastructure represent steps in the right direction. But steps taken at the pace of committee approvals will not close a gap that is widening at the pace of hyperscaler capital expenditure. The window for building meaningful sovereign AI capacity is open. History suggests such windows do not remain so indefinitely.
Peering Ahead
The events of June 2026, the forced withdrawal of Anthropic’s Fable 5 and Mythos 5, the pre-emptive government checkpoint on OpenAI’s GPT-5.6, and Apple’s sweeping price hike on the same day — are not isolated incidents. Read together, they constitute a coherent pattern: the reversal of AI democratisation, visible now at every layer of the stack, from the frontier model to the retail store shelf.
The promise was decentralisation. The reality, increasingly, is the opposite: a handful of corporations owning the physical substrate of intelligence, a single government mediating access to the frontier, memory manufacturers reallocating supply from consumer devices to data centre clusters, and everyone else, the small founder, the independent researcher, the developing-country policymaker, the student who was briefly told the future belonged to her too, left to negotiate the terms of their dependence.
Five corporations own the compute. One government controls the gate. The rest of the world pays the rent, and the rent, as of this week, has gone up.
