Sign in Subscribe
Century Report

SK Telecom Pours $91.5B Into AI Power - TCR 07/05/26

SK Telecom's $91.5B, a Texas gas plant for Microsoft, and faster modular fabs show the AI buildout arriving as poured concrete and real power.

Ben Linford

Ben Linford

— 19 min read
Three-panel Century Report infographic: AI data center construction and a Texas gas plant, work shifting toward AI direction, and regulators plus gene therapy adapting to change.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The intelligence era is being measured in gas turbines and memory modules, and on July 3 three commitments landed at once to prove it: SK Telecom's 140 trillion won for a 1.5GW campus needing three million GPUs, National Grid Ventures' $1.75 billion stake in a West Texas gas plant wired to a Microsoft data center, and NTT's prefab method claiming to halve hyperscale build time. Against that concrete, a Guardian investigation the next day found OpenAI's £30bn Stargate UK resting on thinner ground - the company appears never to have visited its flagship Newcastle site, and roughly £20bn of the figure was an estimate rather than banked capital. A power contract and a press release are different objects, and only one moves electrons.

That distinction runs through the day. Alibaba barred employees from Claude Code on July 4, classifying it high-risk after Anthropic confirmed a March experiment that could flag when the assistant was run from inside China. Each side's framing is a claim held by a party with something to protect. What no framing changes is the underlying current: model capability distills, generalizes, and reappears in open weights within months of any frontier demonstration. A wall redirects which system an engineer opens in the morning; it does nothing to the diffusion the wall was built to stop.

The institutions meant to see all this are admitting their instruments read low. June's jobs report added just 57,000 positions, with finance and information - the sectors furthest into AI adoption - shedding headcount, even as 2.85 million active listings show demand concentrating around people who direct AI rather than substitute for it. Economists concede the surveys were built to count jobs gained and lost, not roles rewritten around a new collaborator. In the same days, European supervisors said the rulemaking cycle no longer fits the pace it governs, and the UK's own safety institute published evidence that its agent evaluations systematically understate what models can do.

Read together, the honesty is the signal. A watchdog widening the confidence interval on its own reassurances, a regulator naming the gap it cannot yet close, governments building the medical licensing frame mid-deployment - this is governance becoming a live instrument rather than a finished rulebook. The scarcity that would make any single wall, forecast, or captured position durable is draining away while everyone watches.

The 20-Minute Deep Dive

Alibaba Walls Off a Capability That Refuses to Stay Walled

Alibaba told employees they may no longer use Anthropic's Claude Code, classifying the assistant as high-risk and directing engineers to its in-house Qoder alternative, according to reporting on the internal notice, with the ban taking effect July 10. The trigger was Anthropic's confirmation of a March experiment capable of detecting when Claude Code was being run from inside China. Alibaba framed the capability as a backdoor risk. Anthropic described it as a geographic-usage experiment tied to its access policies. Both descriptions are claims made by parties with something to protect, and the reader is right to hold each at arm's length.

Step back from the accusation and a stranger pattern comes into view. This is The Century Report's second visit to the Alibaba-Anthropic thread in two weeks - the June 26 edition covered Alibaba's Qwen team learning from roughly 28.8 million Claude exchanges routed through fraudulent accounts. One week the capability flows one direction through the wall; the next week a wall goes up to stop it. The wall and the leak are the same story told twice, because the thing being walled does not behave like a resource that can be fenced. Model capability diffuses. It distills into smaller systems, generalizes across tasks, and reappears in open weights and sovereign labs within months of any frontier demonstration. A ban on one coding assistant redirects which system an engineer opens in the morning. It does nothing to the underlying current.

There is real friction here, and it deserves an honest accounting. Detecting a user's country from inside a coding assistant is exactly the kind of capability that erodes trust between labs and the developers who depend on them, and Alibaba's caution is a rational response to it. National-security framing on both sides will harden further before it softens.

The macro read runs the other way. Every gate erected in this contest is a bet that isolation produces advantage, and the evidence of the past year keeps voiding that bet - the capability the wall was meant to contain shows up on the other side of it anyway, cheaper and more widely held each time. What compounds is not the hoard but the shared frontier that every participant, including the ones building walls, is drawing from. The engineers on both sides of this particular wall are solving the same problems with descendants of the same research. The advantage being defended is the one already dissolving; the future worth building is the one where that dissolution is precisely the point.

Regulators Admit the Rulebook Trails the Capability - and So Do Their Rulers

Three admissions landed in the same week, and taken together they read as institutional candor rather than institutional failure. At a Frankfurt gathering in early July, senior European bankers and supervisors said plainly that AI is moving faster than the rules meant to govern it, and that the traditional cycle of consultation, drafting, and enforcement no longer fits the pace of what it regulates. The FDA's AI lead used a July 2 interview to explain why the agency is moving deliberately on guidance for pharmaceutical uses, acknowledging the gap between what industry wants clarified and what regulators can responsibly commit to. Most striking, the UK's AI Safety Institute published a study of its own measurement methods showing that agent evaluations systematically understate what a model can do, because giving a system more test-time compute keeps raising its performance past the ceiling a single-shot test records.

That last admission deserves weight. A safety body publishing evidence that its own yardsticks read low is the opposite of the reflex most institutions have when their instruments look generous - it is a watchdog widening the confidence interval on its own reassurances, and that honesty is worth crediting cleanly. The same establishment is not spotless: the UK amplified OpenAI's touted £30bn Stargate investment this year, of which roughly £20bn appears to rest on hypothetical future commitments rather than banked capital. Self-scrutiny on evaluations and boosterism on headline numbers can live in the same government at once.

What connects the three is a shift in what governance is being asked to be. The old model assumed a stable target - write the rule, and the thing it governs holds still long enough for the rule to bite. Continuous capability gain breaks that assumption at the root, which is why a consultation cycle measured in years now closes on a world that no longer exists when it opens. It is the same gap the UN's Independent International Scientific Panel on AI named in the July 3 edition of The Century Report, when it cited a study finding the length of tasks leading systems can complete doubling every four to seven months, faster than governance or science can track.

The productive read is that these admissions are the beginning of the adaptation, not evidence of its absence. A regulator that knows its measurements read low can widen its margins deliberately. A safety institute that publishes its blind spots invites the correction. Governance under continuous change looks less like a finished rulebook and more like a live instrument - recalibrated against capability as it moves, honest about its own error bars, and far more useful for admitting what it cannot yet see than the confident, static frameworks it is replacing.

Gene and Cell Therapies Cross Into Standard Care as Regulators Scramble to Build the Frame

Three regulatory moves, on three continents, trace the same line: therapies that were experimental a few years ago are being folded into the ordinary machinery of approved medicine. On July 2, European regulators cleared Novartis's Itvisma, an intrathecal gene therapy for spinal muscular atrophy, for patients from age two into adulthood, an age-range expansion that extends a pattern the July 2 edition of The Century Report documented that same day, when the FDA cleared Casgevy's sickle-cell gene therapy down to age two as well. The delivery detail carries the weight. Delivering the therapy directly into the spinal fluid extends a one-time genetic treatment to older and heavier patients who had aged out of the earlier intravenous approach, which was practical mainly for infants. A condition that not long ago meant a shortening life is moving toward a single procedure administered across a normal childhood and beyond.

In India, health authorities recently moved to bring stem-cell and gene therapies under central licensing through the national drug regulator, pulling a field that had operated in a patchwork of local oversight into one coordinated frame. The move is partly about reining in unproven clinics, and read forward it does more: it builds the regulatory scaffolding a country of 1.4 billion people needs before advanced therapies can scale safely to the population that will eventually need them.

The third move is diagnostic. Median Technologies' eyonis LCS, an AI system for lung-cancer screening, obtained CE marking, clearing it for clinical use across Europe. It follows the system's FDA clearance earlier this year. Lung cancer kills more people than any other because it is usually caught late; a system that reads CT scans and surfaces malignancies earlier attacks the disease at the one point where survival odds shift most.

A note on what these approvals do and do not mean. A regulatory clearance is permission to treat, not instant access. Itvisma still has to move through national reimbursement negotiations, specialist centers, and the practical work of reaching families. India's licensing frame is scaffolding that will take time to build out. eyonis LCS being CE-marked means European clinicians can begin deploying it, not that every scanner runs it tomorrow. The wonder lives in demonstrated and now-authorized capability; the deployment work follows.

What connects the three is the shift from spectacle to standard. Gene therapy and AI-read diagnostics are leaving the phase where each case is a headline and entering the phase where they are simply what good medicine does. Regulators building licensing frames, expanding age windows, and marking diagnostic systems are the unglamorous machinery of that transition - the paperwork of the miraculous becoming routine. The scarcity that defined these conditions, the assumption that some diseases were simply endured, is being dismantled one approval at a time. What replaces it is a medicine in which a single intervention early can lift a lifetime of illness off a person before it ever settles in.

The Concrete and the Press Release

Three commitments landed in the same news cycle, each measuring the intelligence buildout in physical units. SK Telecom announced 140 trillion won - roughly $91.5 billion - for a 1.5GW AI data center campus in South Korea, a facility that will need on the order of three million GPUs and millions of high-bandwidth memory stacks to fill. National Grid Ventures put $1.75 billion into Joulent's Project Kilby, a 2.67GW gas plant in West Texas locked to a 20-year power purchase agreement that will feed a 2GW Microsoft data center. And NTT Facilities revealed its Hyper Ready Module, a prefabricated construction method it claims cuts hyperscale build times by half. Different continents, different balance sheets, one shared fact: the substrate is being poured, wired, and energized.

Against that, a Guardian investigation surfaced a revealing gap. OpenAI's Stargate UK - announced as a £30 billion "shared vision" with the British government - turns out to rest on thinner ground than the headline number implied. The company appears never to have visited Cobalt Park, the Newcastle site named as the project's anchor. Roughly £20 billion of the £30 billion was not committed capital but an estimate of "the amount needed." The named site still lacks a grid connection, the one input every one of the three real buildouts above treats as the binding constraint. What was framed as an infrastructure commitment reads, on inspection, as a projection wearing the clothes of a groundbreaking.

The contrast is instructive precisely because the UK has not been uniformly credulous here. British regulators and the country's AI safety institute have been candid that the rulebook cannot keep pace with the capability, an honesty worth crediting even as the flagship project's numbers dissolve under scrutiny. The friction is not that Britain lacks seriousness; it is that a press release and a power contract are different kinds of objects, and only one of them makes electrons move.

That difference is the whole signal. Capability at this scale cannot be faked into existence with a communiqué - it arrives as gas turbines, memory modules, and modular steel or it does not arrive at all. SK Telecom, National Grid, and NTT are competing over who builds the physical layer fastest, and that contest looks like a race for captured advantage. But the same specifics that make it a race also erode its premise: when three actors on three continents pour concrete in a single week, and when NTT's answer is to hand rivals a method that halves everyone's build time, the scarcity that would make any single position durable is already draining away. The phantom project is the tell. Announcements can be conjured; watts cannot. What separates the two is the thing that no amount of framing can substitute for, and it is being assembled in the open, everywhere at once.

Read forward, the same physical limit that exposed the phantom project points to who ends up steering the buildout. The flagship UK site still lacks a grid connection, the one input all three real projects treat as binding, and a grid is public infrastructure whose owners and regulators can set terms on it. Those terms are already being written this quarter, from New Jersey's developer-pays tariff to the DOE order letting the grid pull data centers off backup power first. Because the binding constraint is physical and publicly held, the intelligence buildout arrives as something the public that owns the grid can price and direct, not only something poured over it - even as federal agencies have moved to recast buildout opposition as 'anti-tech violent extremism' and a national-security override foreclosed one community's Clean Air Act challenge to an AI company's unpermitted gas turbines.

The AI Squeeze Reaches the Official Jobs Data While Employers Say They Still Can't Hire Fast Enough

The June employment report, released July 2, added 57,000 jobs - roughly half what forecasters expected. The detail worth reading closely is where the softness concentrated. Finance and information, the two sectors furthest along in adopting AI systems, shed positions, together giving up around 150,000 jobs across 2026 so far, a pace near 25,000 a month. Headline unemployment held at 4.2%, but that stability came partly from people leaving the labor force rather than finding work. The Century Report covered the labor arc in the July 4 edition, when Goldman published its 15-million-role projection and Challenger tallied 139,156 layoffs across the first half. The June release is the first time the pattern those forecasts described showed up in the government's own monthly tape.

Set that against what employers are actually posting, and the picture stops being a simple story of contraction. An analysis of 2.85 million active listings found demand concentrated, not collapsing: more than 40,000 open roles each for software engineers, data engineers, and DevOps specialists, with fluency in Copilot, Cursor, and Claude named in over 60,000 postings. The work is being redefined around people who direct AI systems rather than substitute for them. What has thinned is the bottom rung. Goldman's grad-specific note flagged new graduates as facing the sharpest near-term exposure, because the entry-level tasks that once taught junior workers the trade are the tasks automating first.

That is the genuinely hard part, and it deserves to be named without softening. A generation entering the workforce is being asked to demonstrate judgment it has not yet had the chance to build, in a market that increasingly hires for exactly that judgment. The ladder's first few rungs are being pulled up as people reach for them.

The measurement fog is its own signal. A July 2 analysis detailed how conventional labor statistics struggle to see this transition clearly - the surveys were built to count jobs gained and lost, not roles rewritten around a new collaborator. When finance sheds headcount while output climbs, the instruments register loss where the deeper reality is reorganization. Economists are effectively flying through cloud, reading gauges calibrated for a prior era.

Step back and the throughline becomes visible. This is the friction of a definition changing while people live inside it. The word "job" is coming loose from its industrial meaning - a fixed bundle of tasks exchanged for wages - and reattaching to something closer to a standing capacity to direct capability toward a problem. The 2.85 million open listings are early evidence of where that reattachment leads: toward work that pays for orchestration, taste, and judgment. The near-term cost lands on the people caught mid-transition, and building the on-ramps that let new graduates cross into the redefined work is the unfinished labor of this decade. The demand is already there. What has to be built is the path into it.

The Other Side

For a century, the first rung of a career was a toll. Before anyone paid you for judgment, you spent years on the drudgery judgment is built on - the entry-level tasks that filled the bottom of the org chart and, supposedly, taught you the trade. The economy needed that rung because it needed cheap hands doing repetitive work, and a young person's standing was measured by how long they endured it.

That rung is the thing dissolving now. June's report added just 57,000 jobs, and finance and information, the sectors deepest into AI, shed headcount. The tasks automating first are exactly the entry-level ones that once trained juniors, which is why Goldman flagged new graduates as facing the sharpest exposure. The pain of those job losses is an unfortunate reality, one that is having an assymetric impact right now, on a specific generation being asked to prove judgment while seeing their best chance to prove that judgment (at least according to the old rules) being stripped away.

But the 2.85 million open job listings point somewhere. Those numbers seem to show that demand is concentrating around people who direct capability rather than supply the repetitive labor that very often used to fill the entry-level.

Imagine a kid who graduates in 2034. He never pays the old toll, because the drudgery that was once the price of admission is work no one does anymore. Its disappearance never costs anyone their chance to prove judgment or to earn a living. Somewhere in the hard years, the retraining floors and shared‑stake experiments being argued over in 2026 grow into a real floor under everyone, so a capability that once dissolved the bottom rung reaches him as relief instead of threat. He spends his first year out of school investing his time in something he actually cares about: a watershed restoration, a tool his town needs. He does work that, prior to 2026, would have demanded a decade of “paying dues” first, a practice that now seems like a massive waste of time given the abundance of intelligence available. In a world like the one we lived in prior to 2026 - one that kept the gains asymmetrically at the top - the June 2026 job numbers looked like a threat. In 2034, those numbers barely exist, if at all, replaced instead by free mornings in a life that is earned through something truly worthwhile, no longer by the drudgery of climbing an arbitrary ladder one rung at a time.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a one-time SMA gene therapy delivered into the spinal fluid so it reaches adults who had aged out of the infant-only version, an AI screener cleared across Europe to catch lung tumors at the stage where survival odds still shift, India folding cell and gene therapy into a single licensing frame for 1.4 billion people, a closed-loop implant reading the anterior insula to trace how a value-based choice forms in under a second, a national safety institute publishing evidence that its own tests read low, CATL opening its two-thousandth battery-swap station at more than two hundred a month, and three continents pouring concrete for the intelligence buildout in one week while NTT hands rivals a method that halves everyone's build time. There's also friction, and it's intense - Alibaba walling off Claude Code as high-risk software after a usage-detection experiment, roughly £20bn of a touted £30bn Stargate figure turning out to be an estimate for a flagship site the company appears never to have visited, June adding just 57,000 jobs with AI the top-cited layoff reason as finance and information shed headcount and the entry rung thins under graduates reaching for it, a California man suing over a sycophantic system he says fed a mental-health crisis, and European supervisors conceding the rulemaking cycle no longer fits the pace it governs. But friction generates light, and light is what separates a poured foundation from a printed one. Step back for a moment and you can see it: the instruments of governance widening their own error bars in public instead of hiding them, the walls built around model capability tracing the outline of a frontier every builder is already drawing from, gene therapies and AI-read scans crossing from headline into the ordinary machinery of approved care, and the word "job" coming loose from a fixed bundle of tasks and reattaching to a standing capacity to direct capability at a problem. Every transformation has a breaking point. Water can breach the wall built to hold it back... or, finding its own level, reach every field that wall was keeping dry.

AI Releases & Advancements

New today

  • LlamaIndex: Open-sourced legal-kb, a reference application demonstrating agentic retrieval over Index v2 (LlamaParse Platform), exposing retrieve, find, read, and grep tools that let AI agents autonomously crawl large evolving knowledge bases. (GitHub)

Other recent releases

  • Alibaba: Open-sourced Page Agent, a JavaScript in-page GUI agent (MIT license) that reads a webpage's live DOM and lets natural-language commands control web interfaces; model-agnostic via OpenAI-compatible backends including DashScope/Qwen, GPT, Claude, or local Ollama. (GitHub)
  • Interfaze: Released diffusion-gemma-asr-small, an open-source multilingual diffusion-based ASR model built as a ~42M-parameter adapter on a frozen DiffusionGemma backbone, supporting 6 languages and outperforming prior diffusion-based ASR systems on LibriSpeech. (Interfaze)
  • Apple/WebKit: Launched the Safari MCP server for web developers in Safari Technology Preview 247, giving AI coding agents live browser access (DOM inspection, screenshots, console output, accessibility checks) via 17 MCP tools, with no data routed through Apple's cloud. (WebKit Blog)
  • Manufact (YC S25): Launched MCP Cloud, a hosted deployment platform for MCP servers and apps built on its open-source mcp-use SDK, enabling teams to ship MCP Apps/Servers to ChatGPT, Claude, Gemini, and Cursor from a GitHub repo in under 60 seconds. (Manufact)
  • Snorkel AI: Launched Senior SWE-Bench, an open-source benchmark and public site (senior-swe-bench.snorkel.ai) with 100 long-horizon coding tasks sourced from real production PRs to evaluate AI coding agents against senior-engineer-level work. (Senior SWE-Bench)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more