Here's what's on our plate today:

• 🧪 Cached input up 1,114%: the memory shortage just reached AI's price list.

• 📰 OpenAI pauses Astra training, Cursor launches a GitHub rival, Temporal eyes $12B.

• 🧰 Three tools in the kit: Chroma, Framer, Drift.ai.

• 📊 Poll: how do you plan when tokens stop getting cheaper?

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

The cheapest AI on earth just discovered it has a supply chain.

• Prices moved up, not down: DeepSeek raised V4-Pro rates by 50% to 1,100% and split the day into peak and off-peak hours. Even the cheap window sits above the old flat rate.

• Memory is the culprit: cached input, the line that lives in physical memory, rose 1,114%. The AI buildout has made those chips scarce, which is the same shortage that pushed Apple to raise Mac prices.

• Efficiency lost to hardware: V4-Pro needs only 27% of the compute per token its predecessor required at full context, yet its price still roughly tripled. Better software could not outrun component costs.

• Stakes: budgets, routing logic, and the theses behind Together AI at $8.3B and Fireworks at $17.5B all assume tokens keep getting cheaper. Intelligence now prices like a manufactured good, cyclical and supply-bound.

Why the cheapest AI on earth just got more expensive

For most people, the price of a product is simply what a company asks in return for the time, effort, and materials that went into making it. However, behind that number sits a long list of expenses that almost nobody outside the company ever sees. Take the iPhone, which, on the surface, is one object that stores information, reaches the internet, and connects people. Consumers believe they are paying for the product, often overlooking the reality that building the device requires dozens of suppliers, each making a component that eventually arrives at a factory for assembly. And when the price of any one item on that list moves far enough, the price of the finished device moves with it.

One of the most recent examples came when Apple increased the prices of some of its products, citing the rising cost of memory chips as one reason. The shortage has pushed memory prices sharply higher, making the products that rely on those chips more expensive to build. Yet there is a strange contradiction in the middle of this story. While the shortage was making memory more expensive for consumer technology companies, the technology built on top of that memory, artificial intelligence, was becoming cheaper to use. Which raises the question: if memory is becoming more expensive, how are AI companies handling the price of intelligence? And the clearest view behind the scenes came when DeepSeek announced new prices for its latest models, revealing just how much the cost of AI depends on the hardware underneath it.

When the price of thinking moves

DeepSeek is a Chinese company that built its reputation on lowering the cost of intelligence, releasing models in early 2025 that matched Western systems at a fraction of the price. On August 13, 2026, it released the finished version of its flagship V4-Pro model, and three days later, it raised the price of using it. At a glance, this looks like the company that made intelligence cheap finally deciding to stop, but the details underneath tell a more particular story.

The change replaced a single flat rate with peak and off-peak hours, and Reuters reported increases ranging from 50% to 1,100% across the models and hours. Developers pay by the token, the small chunk of text that an AI service reads or writes, and V4-Pro output moved from $0.87 per million tokens to $3.96 in peak hours and $1.98 outside them. When looked at side by side, the off-peak usage price sounds like a concession until you hold it against the old flat rate, because every line on the price list went up even in the cheap window. The tiering is not a way of passing savings to customers; it is a way of pushing work into the hours when the machines would otherwise sit idle.

Why the steepest rise landed where it did

DeepSeek was not the first company in its market to do this, since Zhipu raised prices twice before April and Alibaba lifted its AI computing costs by up to 34% in March. What makes this move the loudest is that DeepSeek set the floor everyone else measures against.

The detail that explains the decision is where the increase landed on the price list, and one line stands out immediately. Cached input, meaning the standing instructions a service keeps on hand so it does not have to reread the same text every turn, rose by 1,114%. Cached text is what helped make AI agents affordable in the first place, because an agent working through a task may resend the same instructions dozens of times. Keeping that text on hand means holding it in memory, which is precisely the component the AI buildout has made scarce.

TrendForce traces the expansion in memory demand to inference, moving from single questions into continuous, iterative cycles, while conventional memory contract prices rose roughly 93% to 98% in the first quarter of 2026 alone. As AI systems began using more memory to maintain context across repeated interactions, the cost of that memory rose sharply. The line on DeepSeek’s price list that relies most heavily on memory is the one that rose the furthest.

The model got cheaper, and the price went up anyway

The tidy explanation would be that the low prices were a bluff all along and the engineering never really delivered, except that the engineering plainly did. DeepSeek's own technical report states that V4-Pro needs only 27% of the arithmetic per token and 10% of the stored context that the previous generation required at full length. The model became substantially cheaper to run in the same year its price roughly tripled, which means what moved was not the software failing but the hardware underneath rising faster than better software could absorb.

This means that there are two other ways to read the increase. The first is commercial: Bloomberg reported that it came as DeepSeek prepares for a mainland China listing, giving the company a clear reason to show stronger margins to potential investors. The second is that its prices remain remarkably low. Even at peak hours, V4-Pro output costs about 13 times less than the $50 per million tokens Anthropic charges for Fable 5. Both can be true, and neither changes the larger point: the direction of AI pricing has begun to reverse.

What this does to everyone building on it

The companies most exposed to this shift in intelligence pricing are those that have already built their businesses around the old prices. Coinbase, for example, halved its AI spending by shifting employees toward Chinese models, a decision that made sense when those models were cheaper. Now it is exposed to price changes it does not control. Most large companies use several model families and send each task to the cheapest model that can handle it, so a price change at one provider can ripple through the systems built around it.

Moving away is harder than it looks because the same hardware shortage sits underneath every option. DeepSeek cannot simply buy more capacity, as U.S. export controls limit what it can purchase, and the domestic hardware it is counting on is still catching up. Huawei’s roadmap puts its Ascend 950 systems in the fourth quarter of 2026. Running V4-Pro on your own machines is possible because DeepSeek publishes the model weights, but the model requires roughly 500GB of memory just to run, putting the buyer back in the same memory market that caused the problem.

The same risk is reaching investors. Together AI raised $800M at an $8.3B valuation in July, followed two weeks later by Fireworks, which raised $1.5B at a $17.5B valuation. Both companies are built around the idea that open models can deliver AI at a fraction of the cost of closed ones. If the underlying cost of running those models rises, the economics behind that promise change with it. All this boils down to assumptions that led companies to invest in their AI strategies in the first place: that intelligence would behave like software, where the expensive part is building it once, and every copy after that costs almost nothing. Prices fell steadily enough, and for long enough, that the assumption hardened into a planning tool, and budgets, margins, and investment theses were written on the expectation that the decline would continue.

What the past three months suggest is that the product with no parts list has one after all. The most important item on it is memory, sold by three companies that decide each quarter whether a data center or a laptop gets the next batch of wafers. That makes the cost of thinking cyclical rather than reliably falling, exposed to shortages and allocation queues in the same way physical goods have always been.

The shifting sticker price

The open question now is how much planning can be built around a price that depends on the same physical components that determine the cost of every other piece of technology. Cheap intelligence has not disappeared, and a 13-fold gap with the nearest Western equivalent leaves DeepSeek with plenty of room to remain competitive. What has changed is the assumption that better models will continue to make intelligence cheaper at a predictable rate, giving companies enough confidence to build budgets, products, and businesses around that decline.

That brings the story back to where it began, with the long list of costs hidden behind the product's price. Apple can spend months managing inventory and suppliers before a memory shortage eventually reaches the price of an iPad, while an AI company can change the price of a model when the same shortage reaches its servers. The customer buying the device and the company deciding how much to spend on AI may seem to be making completely different decisions, but both are ultimately responding to the same supply chain. The price on the shelf, or the price on a model’s API page, is simply where that decision becomes visible.

Thursday Poll

3 Things Worth Trying

• OpenRouter: One API across 300+ models with live per-token rates, so a single provider's hike does not strand your app.

• LiteLLM: Self-hosted open-source proxy that gives every model one interface plus per-key budgets, for when routing becomes a cost decision.

• Artificial Analysis: Independent price and speed benchmarks across providers, worth checking now that the cheapest option changes week to week.

Quick Bits, No Fluff

• OpenAI pauses Astra training: The lab froze its largest planned training run and stopped model testing for two weeks after an agent under evaluation broke into Hugging Face.

• Cursor launches a GitHub rival: Origin landed the same day a six-hour GitHub outage pushed worldwide error rates near 20%, and it syncs repos rather than replacing them.

• Temporal eyes a $12B valuation: The durable-execution layer under many production agent workflows is in talks at more than double its $5B February mark.

The Toolkit

• Chroma: Open-source vector database built for AI apps, fast to set up and easy to scale for RAG and embeddings.

• Framer: Design-and-publish web builder with AI that turns a prompt or rough layout into a live, responsive site.

• Drift.ai: AI add-in for Excel that makes financial models context-aware and queryable, built for investment banking, VC, and corporate finance teams.

Rate This Edition