SK Hynix Pours $4B Concrete While AI Code Multiplies

So, What Actually Happened?

Friday, and the thing I keep coming back to is a foundation. Not a metaphor, actual concrete: SK hynix broke ground on a $4 billion plant in an Indiana research park, and S&P put the whole AI build at more than $1.3 trillion by 2027. We scanned 190,000 articles this week so you don't have to. Then I read what developers are actually shipping, and the answer is more code, not more software. Marc Benioff, on the back of a record quarter, called enterprise AI still in its infancy with broad adoption twelve to eighteen months out. I opened the morning reading about supply. I closed it thinking about demand.

The Bottom Line: The spending turned physical this week, and physical spending does not get unwound. The return is still a promise with a date on it, and the date keeps sliding.

The Tracks That Matter

1. SK Hynix Pours Concrete While The Bottleneck Moves Again

SK hynix broke ground on a $4 billion advanced packaging facility in Purdue Research Park, and the word doing the work there is packaging, not chips. Stacking and connecting the memory is where the line backs up now. Days earlier Lam Research opened an Oregon research lab aimed at the same era. And the constraint has already moved past silicon: one market read tracks it from memory to the power gear, where Siemens has gas turbines and heavy transformers booked solid through 2028 to 2030. Every time this industry clears one constraint, the next one has a longer lead time and a bigger physical footprint. You cannot write software that shortens a transformer order book.

Here's what works: Ask your cloud vendor for the physical delivery date behind your 2027 capacity commitment, not the contract signature date. Those are now two different numbers.

2. AI Is Writing More Code And Delivering Less Software

The uncomfortable finding this week: AI-assisted development is producing more code without more delivered software. Generation was never the expensive part. Review, integration, testing and the argument about who is accountable when it breaks, that is where the hours always went, and none of it got cheaper. PagerDuty's engineering team put the question plainly: AI coded it, AI shipped it, who owns it. Salesforce surveyed agentic AI leaders and found preparation beats speed for ROI, which is a polite way of saying the teams who went fastest have the least to show. Australian organisations report the same wall: interest is rising, scaling is not.

Here's what works: Measure your AI tooling on merged pull requests and incidents closed, not on lines generated. Generation is the cheap half of the pipeline.

3. A Hundred AI Companies Ask For Defences Against AI

OpenAI, Anthropic, Google and more than a hundred other companies called for coordinated action against rogue AI. Read the signatory list slowly, because the list is the argument: the firms building the capability are formally asking somebody to build the defence. That request is already a product category. Red Hat's chief executive is pitching agent security and the open-source supply crunch as the partner opportunity of the year, and Filigran shipped XTM One to compress threat response. A joint statement costs nothing. A roadmap change costs something, and only one of those has shown up so far.

Here's what works: Check which of your AI vendors signed. Then ask what shipped differently in the product, not what changed on the policy page.

Quick hits:

• Parallel Web Systems raised $100 million for the machine web. Kleiner Perkins and Index Ventures led a $100 million Series A at a $740 million valuation on Parag Agrawal's premise that agents will use the web far more than humans ever have, mostly in the background where nobody is watching.
• Nvidia is reportedly circling Hugging Face. Talks are reported around $13 billion, which would put the default home of open models inside the company that sells the hardware those models run on.
• Salesforce and Anthropic shipped Claudeforce. CRM data and actions now run inside Claude directly, landing the same week Agentforce annual recurring revenue passed $1.5 billion.

Signal vs. Noise

🟢 Signal: AI governance. People wrote about governance less this week and it mattered more, which is the pattern you see when a topic stops being an opinion column and becomes a standing board item. BDO is now framing it for directors as a resilience question with named crown jewels and concentrated dependencies. Most coverage is still counting policy announcements instead of watching where the agenda item landed.

🔴 Noise: ”agentic AI” as a label. The phrase pulled heavy volume again while its grip on what is actually shipping slipped hard. The real deployments this week filed themselves under advanced packaging, threat response and CRM actions. Not one of them leads with the word agentic.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

SK hynix poured a $4 billion foundation in Indiana, S&P put the AI build past $1.3 trillion by 2027, and a study of developer output found AI is producing more code and not more shipped software, all inside 48 hours.

Three desks, three unrelated stories. The semiconductor wire reads the first as a supply win for American packaging. The ratings desk reads the second as a capex forecast. The developer-tools press reads the third as a tooling problem someone will fix in the next release. Put them on the same morning and the shape is harder to look away from: the supply side has converted into buildings, turbine orders and 2030 delivery slots, while the evidence for the return is still a survey and a promise about next year.

That asymmetry is the whole risk. Software spending is reversible, you cancel the licence at renewal. Concrete is not. When the money becomes a foundation, a groundbreaking and a booked production slot, the industry has committed to a demand curve it has not yet been shown. Somebody is going to be holding capacity they cannot fill.

What changes Monday is small and specific. Look at your own AI budget and split it into the part you could stop paying for in ninety days and the part you could not. If the second number is growing faster than your measured output, you are running the same trade, just at your scale.

By The Numbers

• AI infrastructure investment will exceed $1.3 trillion by 2027 — S&P Global Ratings is now scoring this as a credit story, which means lenders are being asked to price the build, not just admire it.
• SK hynix is putting $4 billion into one advanced packaging plant — packaging, the step nobody outside the industry could name a year ago, is now worth a dedicated American facility.
• AI capital spending is forecast past $650 billion in 2026 and $700 billion in 2027 — more than triple the 2022 level, and the curve is being drawn by six or seven buyers.
• Agentforce annual recurring revenue passed $1.5 billion, up over 240% year over year — the clearest number anyone has published that enterprises will actually pay for agents rather than pilot them.
• Parallel Web Systems closed $100 million at a $740 million valuation — venture money is pricing a web whose main visitors are machines, before anyone has agreed who pays the toll.
• See what's rising across AI and data this week →

Deep Dive: The Best Sounding Empty Room

I once played a club that spent its entire budget on the sound system. Funktion-One rig, room properly tuned, the kind of low end you feel in your sternum before you hear it. Forty people showed up. Best sounding empty room I have ever worked, and the owner could not understand it, because he had bought the best of everything.

The industry is buying the rig

SK hynix is pouring foundations. Lam Research is opening labs. S&P is pricing $1.3 trillion of build. Siemens has turbine slots gone through 2030. This is a genuinely impressive rig and none of it is stupid, because if the demand arrives and you did not build, you are finished.

The door is where the money was made

Meanwhile the delivery numbers sit flat. More code, no more shipped software. Benioff, announcing a record quarter, still says broad adoption is twelve to eighteen months away. Australian firms are stuck between interest and scale. Nobody is short of capability. They are short of throughput.

The tax sits in between

SandboxAQ's Mohammed Aboul-Magd named it: capable agents trapped in separate tools, with a coordination tax eating the productivity they were meant to deliver. Alation's line lands next to it: autonomy without verification is just faster error propagation. That gap between installed capability and delivered work is where this year's budget actually goes.

What Actually Works

1. Split your AI budget by reversibility: what you can stop paying for in ninety days versus what is now concrete, contract or headcount.
2. Change the metric to throughput: merged changes, closed tickets, cases resolved. Anything measuring generation is measuring the cheap half.
3. Name the verification step: for every agent in production, who checks the work and what is the accuracy threshold it has to clear.
4. Count the coordination tax: how many separate tools your agents live in, and how much human time is spent moving context between them.

Great rig, empty floor. The gear was never the hard part; filling the room was.

What's Coming

Packaging Becomes A Word Executives Use

The $4 billion Purdue groundbreaking is the start of a vocabulary shift. Expect packaging capacity and memory allocation to appear in AI supply contracts the way chip allocation did two years ago, and expect the delivery dates attached to them to be quietly longer than anyone planned around.

The Agent Coordination Layer Gets Contested

SandboxAQ putting agents directly into Slack and Teams channels is a bet that the workspace, not the model, is where agent value gets captured. Everyone with a collaboration tool is about to make the same claim, and the winner owns the context your agents run on.

Europe Adds Another Enforcement Desk

Ireland opened its national AI Office this week, which matters more than it reads, given how many US tech firms run their European operations from Dublin. Watch for the first inquiry it opens; that will tell you what it actually intends to police.

For Your Team

Monday's meeting prompt: ”Split our AI spend into what we could stop paying for in ninety days and what we could not. Is the second number growing faster than anything we can measure coming out the other end?”

Share-worthy stat: S&P now expects AI infrastructure investment to pass $1.3 trillion by 2027, while the best available read on developer output says AI is producing more code and not more delivered software. Both of those were published this week.

Go deeper: Track where AI infrastructure money is landing in real time →

The Track of the Day

Every team we talk to has capable agents trapped in separate tools, and a coordination tax eating the productivity those agents were supposed to deliver.
Mohammed Aboul-Magd, SandboxAQ

That is the sentence I would put in front of anyone approving next year's AI budget. The capability is bought. The tax is what you have not costed yet.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: August 28, 2026 | Curated by Yves Mulkers @ Ins7ghts