AI is making hardware expensive. Software is next.

Three months ago, a Mac Studio M4 Max with 128GB unified memory and 2TB storage was available in India for ~₹4.24 lacs.

Today, the 64GB variant itself is closer to ~₹5 lacs.

More money. Half the memory.

That looks like an Apple pricing quirk until you sit with it for a minute. Many companies are still budgeting for AI as if the main cost will be model subscriptions. That may be too narrow. The first surprise could show up in more ordinary places: laptop refreshes, storage upgrades, local compute, and software plans that used to feel generous.

AI still feels cheap on the surface. A few thousand rupees a month can buy writing, coding, research, summarization, image generation, and workflow automation.

Underneath, the cost curve is not nearly as friendly.

That gap can hold for a while. It cannot hold forever.

For the last two years, teams have added AI into product, engineering, support, sales, and marketing without feeling much of the infrastructure bill. That bill is now moving through the stack.

First hardware.

Then software.

Then the way companies design workflows.

This week in AI News

A few fresh developments from the June 25–July 2 window point in the same direction:

• OpenAI’s GPT-5.6 preview turns model choice into product strategy. OpenAI’s Sol, Terra, and Luna split makes cost, speed, and capability tradeoffs explicit.
• Adobe’s Topaz deal shows creative AI moving into the core workflow. The value sits less in new generation tricks and more in AI becoming part of professional creative production.
• Onsemi’s Synaptics deal shows edge AI becoming a stack decision. Physical AI depends on sensing, power, connectivity, and compute moving together.

This is bigger than one Apple configuration

It would be easy to dismiss the Mac Studio example as one expensive machine in one market.

The more useful question is different.

What else in the company assumes memory, storage, model access, and AI usage will keep getting cheaper?

That is the tension running through the current AI boom. Companies are redesigning work around intelligence that still feels subsidized. Developers use coding agents as daily infrastructure. Support teams route more work through copilots. Product teams depend on models for research, synthesis, prioritization, and documentation.

That may be the right move.

The mistake is acting as if the price of that dependency is settled.

Memory became strategic infrastructure

Memory used to be the boring part of computing.

Useful. Cyclical. Invisible until you ran out of it.

AI changed that.

Modern AI systems need more than GPUs. They need high-bandwidth memory, server DRAM, enterprise SSDs, storage capacity, networking, power, cooling, and long-term supply commitments. The accelerator gets the attention. The rest of the stack sets the constraint.

TrendForce’s recent 1Q26 memory outlook and 2Q26 AI server demand report show how violent the shift has become:

Those are not normal background fluctuations. They are supply-chain stress signals from AI demand spilling into the broader computing market.

A laptop buyer does not think they are bidding against a model lab. A phone maker does not think it is competing with an AI data center. An enterprise IT team buying developer machines does not think it is part of the same market as frontier-model inference.

The supply chain does not care about those categories.

When AI demand pulls memory and storage into servers, everyone else feels it later.

Apple is where the signal becomes visible

Apple did not create this problem. Apple is just where the signal becomes easy to see.

Reuters reported via TechCentral that Apple raised prices on several MacBook and iPad configurations as memory and storage costs surged. The examples matter because they turn an abstract component shortage into prices customers can actually see.

The pattern matters more than the exact SKU.

If Apple, with its scale and supply-chain discipline, starts passing through memory costs, smaller device makers have less room to absorb the shock.

Reuters separately described an acute global memory-chip shortage, with AI and consumer-electronics companies competing for limited supply. It later reported that rising memory prices were already weighing on the outlook for smartphones and PCs. The pressure has moved out of theory and into the device market.

For product leaders, this changes the cost base of everyday work:

• developer workstations
• local AI machines
• employee laptops
• storage-heavy creative workflows
• edge AI devices
• enterprise hardware refreshes
• product experiences that assume more memory at the edge

The first visible AI cost shock is not a subscription invoice.

It is a more expensive machine.

The market noticed before customers did

The stock market saw the memory story before most consumers did.

The cleanest example is SanDisk, using Nasdaq historical data as the market reference.

The exact peak is not the point. The repricing is.

Memory started being treated less like a sleepy component cycle and more like AI infrastructure.

A fair objection: memory is cyclical. Prices overshoot. Supply responds. A single stock chart should not become a permanent forecast.

True.

But the product lesson does not need memory prices to rise forever. It only needs a simpler idea: AI value will not stay trapped inside AI labs. It will leak into the layers that make AI possible.

That should change how teams think about adoption.

Software will not be immune

Hardware shows the pressure first because component costs are concrete. A bill of materials goes up. A product price changes. Customers notice.

Software can hide the pressure longer.

That makes the next phase more dangerous.

ChatGPT, Claude, Claude Code, Cursor, Perplexity, Gemini, GitHub Copilot, and similar tools still feel underpriced relative to what they provide. For ~$20/month, an individual user can access capabilities that would have sounded impossible a few years ago.

But ~$20/month is an adoption-era price point, not a law of nature.

The infrastructure underneath is expensive. Reuters reported that OpenAI was targeting ~$600 billion in compute spending through 2030. Microsoft has separately disclosed that OpenAI contracted to purchase an incremental $250 billion of Azure services. The exact long-term economics are still uncertain, but the direction is clear: serving intelligence at global scale takes serious capital.

Software packaging is already adapting.

GitHub makes the same shift visible from the developer side: as Copilot supports more agentic workflows, pricing moves closer to usage.

Operator takeaway: The product risk is not that AI gets expensive. The risk is that teams redesign workflows around cheap AI before they know which parts of the workflow actually deserve frontier-model economics.

Software pricing will probably not move through one dramatic price hike everywhere. It will be quieter:

• usage caps
• premium model tiers
• paid overages
• slower default models
• enterprise-only features
• agentic workflow limits
• model routing between cheap and expensive intelligence
• sharper separation between casual use and production use

The product may still say "AI assistant."

The economics underneath will become much more segmented.

AI still feels like software. Its cost structure is starting to behave like infrastructure.

The workflow gets repriced after the habit forms

The bigger risk is organizational muscle memory. Teams change how they work first, then discover later what the new workflow actually costs.

Picture a product and engineering team that has quietly made AI part of the workday:

• engineers use coding agents for scaffolding, tests, refactors, and code review
• PMs use frontier models for customer-call synthesis, PRD drafts, roadmap tradeoff memos, and competitive research
• support uses copilots to summarize tickets and draft responses
• design uses generative tools for variation, cleanup, and handoff
• sales and customer success use AI to prep accounts and summarize meetings

At first, this looks like pure productivity upside.

Then the packaging changes. The best model moves into a higher tier. The coding agent gets usage caps. The support workflow needs enterprise controls. The design tool prices heavy generation differently. Local AI machines cost more during refresh. Finance sees five different AI bills, but no one can map them to business outcomes.

By then, the usefulness debate is over. The harder question is whether the company knows which workflows now depend on subsidized intelligence.

That is a product problem, not a finance cleanup exercise.

Pricing, latency, usage limits, fallback modes, and model routing now belong in product architecture. Teams that decide them late will either damage margins or disappoint users when limits become visible.

Product leaders need cost-resilient AI adoption

The next phase of AI adoption can stay ambitious. It needs to be less naive.

Product and business leaders need to stop asking only, "How do we get more people using AI?" They also need to ask, "Which parts of our operating model still make sense when AI is priced closer to its underlying cost?"

A practical audit can start with five checks:

That does not make AI adoption a mistake. It makes lazy adoption expensive.

Use AI aggressively where the gain is real. Build with it. Redesign workflows around it. But do it with a clear view of cost, dependency, fallback, and ownership.

The subsidy phase is ending

The first AI shock was capability.

People discovered that models could write, code, summarize, search, reason, generate, and operate tools.

The second AI shock is cost.

Hardware is already showing the signal. Memory prices surged. Storage became strategic. Consumer devices got more expensive. A memory stock briefly moved like a frontier AI company. The supposedly boring layers of computing stopped looking boring.

Software comes next. AI tools will remain useful. The harder question is whether companies are building workflows that still make sense when pricing reflects the infrastructure underneath.

Teams that treat today’s AI pricing as permanent may be building on a subsidy without naming it.