Apple's Russian
Gas Problem.

Apple outsourced its AI future to the one company building everything that competes with iPhone. The Russian gas analogy fits.

I build software in Bartlesville, Oklahoma. Most of what I do lately is a local-first product called Capsule — a permanent memory layer for AI. I'm not a Silicon Valley analyst. I pick a stack on Monday and ship something on Friday, and the AI landscape looks different from where I'm sitting than it does in most of the analysis.

The Russian gas analogy

Apple's AI strategy right now is to buy winter gas from the Russians. They've outsourced the frontier-model layer of their entire AI future to Google, the one company that is also building the phone, the watch, the glasses, the laptop, the search engine, the browser, the maps, and the email client that compete most directly with Apple's. Every dollar Apple sends Google funds the team building Gemini Intelligence on Pixel.

We saw how that worked out for Europe. The dependency looks fine right up until the moment it doesn't, and by then you've atrophied the muscles you'd need to do it yourself.

The original reporting suggests Apple was close to a deal with Anthropic and balked at the price. Then it was OpenAI for ten minutes. Then Google teleported their best salesperson to Cupertino and that was that. I'd bet real money the salesperson's marching orders flipped the moment the contract was signed — from win Apple to do not lose Apple under any circumstance. That's a sticky arrangement, and it explains why even an obvious strategic mistake can survive longer than it should.

Phones — Google won

Google had I/O 2026 last week. They shipped Gemini 3.5 Flash. They announced Gemini Omni, a multimodal model that generates any output from any input starting with video. They introduced Gemini Spark, a general-purpose agent that takes action on your behalf across connected apps. They rolled out a new design language called Neural Expressive across Android, iOS, and web. They committed Gemini Intelligence to watches, cars, glasses, and laptops, starting with Samsung and Pixel phones this summer.

Apple shipped Apple Intelligence last year and has been refining it incrementally since. WWDC arrives June 8, and the leaked previews suggest more of the same — a redesigned standalone Siri app, an app intents API, better on-device models. Apple also recently agreed to pay $250 million to settle a class action over Siri features they marketed but didn't ship on time. iPhone sales won't crater; they never do. But the narrative gap, already wide, will widen.

Two details from the latest reporting deserve specific attention. First, the new Siri app reportedly leads with auto-deleting conversations — 30 days, one year, or forever — pitched as a privacy advantage over ChatGPT, Gemini, and Claude. That's a strange flagship feature at a moment when the rest of the industry is racing in the opposite direction. AI is becoming a planning partner, not a search query. The value compounds with memory. Forgetting on a timer is what a 2014 chatbot does. It's not what a 2026 AI assistant should be marketing as a differentiator.

Second, the same reporting confirms that the new Siri will be powered by Google's Gemini — but Apple won't emphasize this because Google's reputation as an ad-funded data company conflicts with Apple's privacy branding. Publicly distancing yourself from your own model supplier is not a sustainable position. It's the visible tell of a decision that didn't fit and that no amount of marketing will smooth over.

Siri, AI aside, is in a sad state. Setting a timer and playing music is about the ceiling of reliability. The pieces to fix it have existed for years — local LLMs running on a Mac Mini do better intent parsing than Siri does on a $1,200 phone. The problem isn't technical. It's organizational. Siri sits between the OS team, the services team, and the AI team, and nobody fully owns making it good.

Or look at autocorrect. It's 2026 and iOS autocorrect is genuinely broken. Not slightly behind. Broken. It corrects words in isolation, makes wrong substitutions confidently, and changes simple things like blend into Belinda. Other platforms have flaws, but none of them are this. Whatever's responsible — AI, no AI, organizational drift — the result is a daily, accumulating piece of evidence that Apple is no longer where reliable basic features come from.

To Google's credit, they were caught off guard by ChatGPT just like everyone else. Sundar Pichai has said so on the record. The difference is they got busy fast — pulled their best people, restructured the org, and started shipping. Apple was caught off guard too. They are not, by any visible measure, shipping at the same pace. Three and a half years in, the Apple answer is still “we'll show you at WWDC.”

Desktops — users won, and they're expanding

On the Mac, Apple isn't behind in AI. They're irrelevant in AI. There's a difference.

For fifteen years, power users have been quietly assembling their own desktop intelligence layer out of Raycast, Alfred, Hazel, Keyboard Maestro, Shortcuts, BetterTouchTool, Obsidian, DEVONthink, and a hundred smaller tools. These setups are personal in a way an OS-level feature can never be. When you've spent six months tuning your Raycast extensions and your Obsidian vault and your folder action scripts, the tooling becomes yours. That's authorship. People don't give up authorship.

So if Apple ships “AI triages your inbox” in macOS 27, it will land with a shrug. People who care about that problem are already using Superhuman, Shortwave, or piping their mail through Claude with custom prompts. People who don't care won't notice. There's no middle market on the Mac for AI features Apple ships from the top down.

Apple built the best local-AI hardware on the planet. Unified memory, neural engine, ridiculous performance-per-watt. And the people getting the most out of it are running Ollama, MLX, LM Studio, and bespoke Python — none of which Apple ships, supports, or particularly acknowledges. The M-chip is doing the heavy lifting for a software ecosystem Apple isn't part of.

Here's the part the analysis usually misses: users aren't just winning the desktop. They're expanding their footprint. Local-first is going mainstream and most people don't fully realize it yet. The drivers are simple. People are tired of paying monthly forever for tools that quietly degrade. Tired of their data living on someone else's server. Tired of vendors changing the terms when convenient. Tired of waking up to find their workflow has been “improved” overnight. Obsidian, Logseq, Anytype, iA Writer, Reflect — the list keeps growing. Tools that run on your machine, store data in formats you own, and don't require a login to work.

The AI wave accelerates this, not slows it. The moment your second brain runs on someone else's server, it isn't yours anymore. And once people get a taste of running real models on their own hardware — fast, private, no monthly fee, no upstream policy change that breaks their workflow — the genie doesn't go back in the bottle.

My personal bet, and I'm putting time and money on this, is that the next two to three years bring a growing central retrievable repository of personal intelligence sitting on your own machine, queryable by whatever model you trust — even from your phone if your business allows it. The veterinary practice with several doctors who needs every chart, every case discussion, every conversation findable across the practice. The CEO retiring who wants the company's institutional knowledge captured before it walks out the door. The small company doing a handoff to the next generation. Verticals with hard privacy requirements — legal, healthcare, financial services — where the data legally can't leave the building. None of these have good off-the-shelf solutions today. Most users don't yet know they need them. They will.

The framing shift underneath all of this is the one nobody's named yet: AI is moving from a question-and-answer Google-search lookalike to a useful, reliable partner tool that lives inside whatever you do. That's a much bigger market than search. And it's a market where local-first has a structural advantage, not a disadvantage.

Wearables — the one that matters

Wearables are the only growth market left in consumer computing, and Apple is positioned to lose them worst.

Phones are saturated. Desktops are saturated. Smart glasses are where the next platform shift probably happens, because they're the first form factor that makes ambient AI actually ambient — always available, hands-free, visual context built in. Whoever owns that surface owns the next decade the way whoever owned mobile owned the last one.

Wearables play to Google's strengths and against Apple's. Glasses need real-time information lookup, search, location intelligence, on-device vision ML, and a model that reasons across all your data. Google has been building the AI stack underneath Android, Search, and Maps for a long time, well before anyone called it AI. Apple's strengths — industrial design, retail, brand — matter less when the moat is the AI layer feeding the glasses. The glasses themselves are just a screen and some cameras.

OpenAI is stalled here. The Jony Ive partnership was announced with enormous fanfare and then went quiet, which usually means harder than expected. Apple has AirPods and the Apple Watch — two of the most successful wearables ever shipped — and somehow neither is an AI story. The Vision Pro is a $3,500 demo unit nobody talks about anymore. Their glasses product has been “next year” for four years running.

AirPods with cameras have been rumored for a while and would be a credible Apple play. They leverage the strongest wearable category Apple already owns and skip the harder problem of convincing people to wear glasses every day. But a credible rumor isn't a ship date. Google ships this fall.

What Apple should do

Drop Google. Reconnect with Anthropic. Pay the price this time. Anthropic is the one frontier lab whose stated values match Apple's stated values — safety-forward, privacy-conscious — and unlike Google, they aren't shipping a competing phone, watch, glasses, search engine, or anything else. Every dollar Apple sends Anthropic funds models, not hardware that competes with iPhones. That's the structural difference, and it's huge.

The price might not be what it was, either. The reported Anthropic deal Apple balked at probably included API access and inference tokens bundled in — that's where most of the cost lives at this scale. But Apple already runs a BYO pattern today: users sign into their ChatGPT account so Siri can use it. Run that pattern wider. Let users sign into their Anthropic account, their OpenAI account, their Perplexity account. The tokens move through user accounts, not Apple's. Suddenly the conversation with Anthropic is about orchestration, partnership, and engineering rather than gross inference cost. Much cheaper second coffee.

Apple has also had time to watch Anthropic in action since the first round of conversations. The principles are real, the safety story is credible, and the model quality is genuinely competitive. The Anthropic of 2026 is a stronger partner than the Anthropic of 2024.

Beyond the model, add full BYO. Let users plug in whatever they want for whatever task. Claude for writing, a local model for private stuff, ChatGPT if that's their preference, Gemini if they really insist. Apple's job becomes the orchestration layer — permissions, on-device context, app intents, the secure enclave for personal data. If there's a complaint about the model, it's on the model. If there's a complaint about the experience, that's Apple's lane, and it's the lane they're good at.

On search, the harder truth is the BYO transition is already underway and Apple isn't part of it. Most of my day, Claude is open. None of my tabs are Google search. People aren't opening Safari to look something up anymore — they're opening Claude or ChatGPT or Perplexity. Which means Google's leverage with Apple isn't as strong as the $20 billion suggests. That money buys default position in a browser people increasingly bypass for actual information lookup. Google's real defense is to blend Gemini into Google search itself and keep paying Apple to stay in front of users — banking on the gravitational pull of default behavior. That works for a while. It doesn't work forever.

Why none of this will happen

Google pays Apple roughly $20 billion a year to be the default search engine across Apple surfaces. That deal drives nearly half of Google's search traffic. The September 2025 antitrust ruling let the payments continue with conditions, and now Apple can renegotiate annually and even introduce a choice screen. The leverage is sitting right there on Apple's side of the table.

But the Gemini-as-AI-supplier deal almost certainly isn't a clean separate transaction. It's entangled with the search deal. Google's account team didn't win the AI placement on its merits. They won it by bundling — keep us as default search, let us be your AI brain, we'll make the overall numbers work for you. That's how billion-dollar accounts operate. They don't defend the line item. They expand the relationship until cutting any single piece becomes unthinkable.

The spreadsheet problem is always real. Apple's CFO has fiduciary duties. $20 billion in nearly-pure-profit services revenue isn't a number you walk away from on a hunch. But the spreadsheet that protects the quarter is not the same spreadsheet that protects the decade. Wearables are the decade. Losing wearables to Google because you wouldn't untangle from a deal that made sense months ago is a much more expensive line item than the one the CFO is currently optimizing for.

Apple's path to dropping Gemini isn't switch AI providers. It's unwind a deeply entangled multi-product relationship with the only company paying us $20 billion a year in nearly-pure-profit services revenue. That's not a strategy problem. That's not even a leadership problem. That's a willingness-to-eat-the-quarter problem. And in the Tim Cook era, eating the quarter is the one thing Apple has trained itself never to do.

Apple's done this before

Apple owned a software company called Claris. You've probably never heard of it. That's the point.

The product underneath it is FileMaker, which I've been writing for thirty years and which quietly powers small businesses, schools, dental practices, and law offices across the country. Apple bought Claris in the late eighties, spun it out, renamed it FileMaker Inc, and then rebranded back to Claris in 2019. Three decades of corporate indecision about what to do with a perfectly good asset. The product still ships. It still has loyal customers. It's also a fraction of what it could have been if Apple had ever decided what they actually wanted from it.

Or look at how Apple does new categories. When cloud was emerging, Apple shipped MobileMe in 2008. It was bad — broken sync, missing features, embarrassing reviews. Steve Jobs reportedly gathered the team and asked them what MobileMe was supposed to do. When someone answered, he said: “So why the f___ doesn't it do that?” He replaced the leader on the spot. It took until 2011 to relaunch as iCloud, and several years after that for iCloud to be reliable.

With Steve there was a forcing function. Without him, the same pattern still happens, just with slower correction. AI looks a lot like MobileMe right now — except this time there's no Steve to bang the table.

The sundowning

If the CFO wins this argument, and the CFO usually wins, Apple doesn't die. The brand is too strong. The install base too sticky. But every company has a window of relevance, and the window doesn't close with a bang. It closes with a series of decisions that each look defensible — protect the quarter, honor the contract, don't risk the relationship — and collectively add up to a slow sundowning. IBM didn't die. Nokia didn't die. They just stopped being where anything important happened.

That's the risk. Not collapse, but irrelevance. In 2030 the iPhone is still selling fine and nobody under thirty has one, because the people they follow all wear something else. Google owns the assistant. Google owns the glasses. Google owns the search. And Siri, bless her, gives you great recipes to bake.

For the rest of us — the builders, the indie devs, the local-first crowd — this is good news. The desktop window is wide open. The personal-knowledge layer is underbuilt by everyone. The big platforms are sleepwalking. We get to keep building.

I'll be in Bartlesville, shipping Capsule, watching WWDC. Hoping Apple proves me wrong. Rooting for it. Not waiting for it.