> daily_signal(2026_07_11)

Apple handed Broadcom more than $30 billion for chips and called the order Wi-Fi and Bluetooth parts.

PickBits Daily Signal · Saturday, July 11, 2026

By Mark Pickering · 9 min read · July 11, 2026

// tl;dr

Apple committed more than $30 billion to Broadcom, its largest US manufacturing commitment to date, and publicly called it a Wi-Fi and Bluetooth chip order. Broadcom will make more than 15 billion US-made chips for Apple at its Colorado site, and Apple is putting $1.5 billion into Broadcom's Fort Collins plant. Apple will not detail what else is in there; the reporting says part of the work is Apple's first dedicated AI server chip.
Meta priced its Muse Spark 1.1 coding API below OpenAI's and Anthropic's comparable rates, aiming straight at the tokens developers spend the most on. The price war itself is not new. It has been running since DeepSeek made a 75 percent cut permanent in May, and in June the labs split openly: OpenAI pushed effective cost down while Anthropic raised prices with Fable 5.
Patreon switched on Cloudflare's crawler blocking so AI training bots cannot scrape the members-only posts creators charge for. It runs on Cloudflare's July 6 upgrade from a single AI-bot switch to per-purpose controls that separate search indexing, model training, and agent traffic. Patreon set it at the platform level, so every creator page is covered without anyone touching a setting.
Anthropic launched in-house AI drug-discovery programs and a bioscience product, Claude Science, aimed at diseases the pharmaceutical industry has judged too unprofitable to chase. The pitch is that AI cuts the cost of the early stages enough to make neglected and rare conditions worth researching. It'll take years to know if that's true: more than 200 AI-discovered drugs are already in trials, and none have been approved.

Back in April we wrote that Apple was paying Google around a billion dollars a year for the AI behind Siri, on top of the search deal, because it couldn't build a good enough model in time. That's still true this morning. What's changed is that Apple has now handed Broadcom more than thirty billion dollars for silicon, and it won't quite say what the silicon does. Apple has given up on winning the model race, and it's making very sure it owns the hardware underneath whoever does. Meta spent the same week going the other direction, cutting its coding API in under OpenAI and Anthropic, which tells you exactly what Meta reckons the model layer is worth now. Patreon dropped a layer lower and locked the AI crawlers out of the work its creators get paid for. Anthropic did the one thing here that isn't about market share at all, and started drug programs for diseases that have never made anybody a dollar.

Today: Apple committed more than $30 billion to Broadcom in its largest US manufacturing deal, Meta cut its coding API under OpenAI and Anthropic, Patreon switched on Cloudflare's blocking to keep training crawlers off creators' paid posts, and Anthropic started drug programs for the diseases the industry has priced out.

1. Apple spent more than $30 billion with Broadcom and called it a Wi-Fi chip order.

Nobody spends thirty billion dollars on Bluetooth.

Apple expanded its relationship with Broadcom to the tune of more than $30 billion, its largest US manufacturing commitment to date. Broadcom will produce more than 15 billion US-made chips for Apple at its Colorado site, covering custom silicon across multiple generations of Apple devices, and Apple is putting $1.5 billion directly into Broadcom's Fort Collins plant. The headline number folds into the $600 billion Apple has already pledged to spend domestically, and Tim Cook called it another step toward an end-to-end silicon supply chain in the US. Broadcom's stock closed up almost 6 percent, its sharpest rally since February.

Now ask Apple what the chips do. Apple says they are components that help its devices connect to Wi-Fi and Bluetooth, and it will not give you a word more. Bloomberg's reporting is that part of the work is tied to Apple's first dedicated AI server chip, the silicon that would run Apple Intelligence on Apple's own Private Cloud Compute machines. Thirty billion dollars is an awful lot of Bluetooth. None of it is mysterious once you remember what Apple has been doing all year: in April the company confirmed it pays Google about $1 billion a year for the Gemini access behind the Siri requests it can't handle on-device, a contract running to at least 2029. Apple has quietly conceded the model race and gone to buy the ground the models stand on. And look who's on the other side of the signature. In June, Broadcom unveiled Jalapeño, OpenAI's first custom inference chip, built for the servers answering ChatGPT requests. Same contractor, three weeks apart. Broadcom has quietly become the shop you call once you have decided you would rather not be Nvidia's customer.

techinasia July 9 2026 report Apple signs chip supply agreement worth more than 30 billion dollars with Broadcom its largest US manufacturing commitment covering more than 15 billion US-made chips at Broadcom's Colorado site with 1.5 billion dollars into the Fort Collins plant Apple describes the components as Wi-Fi and Bluetooth while Bloomberg reports part of the work is Apple's first dedicated AI server chip — techinasia.com · July 9, 2026

Why this matters: If any part of your work runs on AI, the price you pay for it is being set right now by chips you will never see, in deals like this one. For years every AI workload ran on more or less the same Nvidia part, and that is ending, because the biggest buyers worked out that renting general-purpose silicon sized for somebody else's model is a bad trade. What replaces it is a pile of custom chips, each one cheap for the company that owns it and useless to everybody else. Action this week: Open your cloud or AI provider's roadmap and find out which accelerator your workloads actually land on today, because almost nobody has checked. If a pipeline is welded to one vendor's toolchain, loosen it now while loosening is still cheap, so get a portable runtime in front of it before the hardware underneath you changes without asking. Follow this one, because the custom-silicon build-out moves in thirty-billion-dollar steps and we'll keep tracking who signs next.

techinasia.com: Apple signed a $30B+ chip deal with Broadcom (July 9, 2026)
techinasia.com: Broadcom and Apple extend custom chip deal to 2031 (July 2026)

2. Meta cut its coding API under OpenAI and Anthropic. The price war is older than today.

Somebody is eating this bill, and it is worth working out who.

Meta released Muse Spark 1.1 and set its API pricing below OpenAI's and Anthropic's comparable coding-model rates, with the 1.1 rollout landing on July 10. The target is not subtle. Code generation and agentic coding are the highest-volume, highest-value tokens in enterprise AI, and they are the segment Claude and OpenAI's models have led. Muse Spark is Meta's coding line, and 1.1 is an incremental release whose news is the number on the invoice, not a new capability tier.

Plenty of people are calling this the opening shot, and it isn't. We have been tracking this fight since May 23, when DeepSeek made its 75 percent V4-Pro discount permanent and put output tokens at $0.87 per million against roughly $30 for GPT-5.5, about 34 times cheaper. By June the frontier labs had split in public: OpenAI pushed effective cost down with flexible rate-limit resets on Codex, and Anthropic went the other way, charging twice as much for Fable 5 in exchange for about 5.7 percent more performance. So Meta has not started anything. It has dragged DeepSeek's logic into the one workload that actually pays the bills. My problem with the discount is the arithmetic underneath it. Coding is the most expensive inference there is, because agentic coding means long context, long output, and many turns, and it just got the biggest discount on the market. That cost does not go away because a pricing page changed, so somebody is absorbing it. Anthropic filed confidentially for an IPO in June, which means we may finally get to read exactly who is absorbing what, in a document with legal consequences for lying.

the-decoder July 9 2026 report Meta Muse Spark 1.1 API pricing squeezes OpenAI and Anthropic as the AI price war heats up coding model rates undercut competitors — the-decoder.com · July 9, 2026

Why this matters: If your team pays for an AI coding assistant, that bill is about to get cheaper, and the cheaper it gets the more carefully you should read the terms. A promotional rate is not a contract, and the last few rounds of this have already shown what happens when the unit economics catch up: the headline price stays, and the cost comes back as a surcharge, a rate limit, or a quietly worse model behind the same name. Action this week: Put your current provider's rates and Meta's Muse Spark 1.1 rates side by side on your real workloads, measured as cost per finished task rather than cost per million tokens, because a cheaper token that needs three more turns is not cheaper. Before you migrate anything, check three things in the fine print: the rate limits, the data-retention and training policy, and whether the price is contractual or introductory. And pilot behind an abstraction layer so a price reversal costs you a config change instead of a rewrite. Send this to whoever signs off on your engineering tooling budget, because they are about to be pitched a discount and they should know what the last four rounds of this looked like.

the-decoder.com: Meta's Muse Spark 1.1 API pricing squeezes OpenAI and Anthropic as the AI price war heats up (July 9, 2026)
theverge.com: Meta's Muse Spark model API (July 2026)
techinasia.com: Meta rolls out Muse Spark 1.1 AI update (July 10, 2026)

3. Patreon locked the training crawlers out of work that creators get paid for.

The courts are slow and the contracts are narrow. This one is a checkbox.

Patreon turned on Cloudflare's AI-crawler blocking so that bots collecting training data cannot scrape the members-only posts creators sell: the art, writing, music, and video that subscribers pay for. It runs on the granular bot controls Cloudflare shipped on July 6, which replaced a single all-or-nothing AI-bot switch with per-purpose settings that sort crawlers into search indexing, model training, and autonomous agent traffic, so a site can keep the search bots and turn away the training scrapers. Patreon set it at the platform level, which means every creator page it hosts is covered without a single creator changing a setting.

We have watched the consent fight try every other route this year, and they are all slow. In May, Amnesty International's Unlawful by Design report called bulk training pipelines a mass invasion of privacy. In June, Hollywood's actors got AI consent and pay written into their contract, which works beautifully if you happen to have a union. Later that month a federal appeals court heard, for the first time, whether training on copyrighted work is fair use, and last week a hundred authors who opted out of Anthropic's book-piracy settlement started suing individually. Every one of those routes will take years, and Patreon took an afternoon. Cloudflare's own June data was that AI bots now generate more web traffic than humans, and its CEO said the web's future is "pay to crawl." Patreon went blunter than pay. If you build models, read that carefully, because nobody is going to litigate the free pile out of existence. It is simply getting switched off, one platform at a time, by an infrastructure company with a checkbox.

404 Media July 9 2026 report Patreon partners with Cloudflare to block AI crawlers from scraping creators members-only paid posts for model training data using new granular bot controls separating search training and agent crawlers — 404media.co · July 9, 2026

Why this matters: If you publish anything online that people pay you for, the door on your work being taken for free just got a lock, and it only clicks shut on the platforms that turned it on. Patreon covered its creators at the platform level. Your own site did not get that for free. The useful part of Cloudflare's change is that blocking training scrapers no longer means blocking Google too, which was the reason most people left the old switch alone. Action this week: Open your Cloudflare dashboard, or ask whoever runs your hosting, and set the AI crawler controls deliberately: block the training category, keep search allowed, and then actually verify it took effect rather than trusting the toggle. While you are in there, check your robots.txt and any AI opt-out signals so they say the same thing. If you build or buy systems that ingest web data, you have the opposite job: confirm your crawlers honor per-purpose directives, and start writing licensed sources into next year's budget, because "we didn't know it was blocked" stops being a defense the moment the block is a standard setting. Send this to a creator you know, because most of them still assume this is somebody else's job.

404media.co: Patreon blocks crawlers from stealing creators' work for AI training (July 9, 2026)
the-decoder.com: Cloudflare replaces its blanket AI bot block with granular controls for search, training, and agent crawlers (July 6, 2026)

4. Anthropic started its own drug programs, aimed at the diseases the industry decided were not worth the money.

Those drugs don't exist because a spreadsheet somewhere said no.

Anthropic said it will develop its own drugs, standing up in-house AI drug-discovery programs and a bioscience product called Claude Science to run them. The target it named out loud is the interesting part: the diseases Big Pharma considers unprofitable, meaning the neglected and rare conditions where the commercial return never justified the research spend, rather than the blockbuster markets the industry crowds into. The models get pointed at the early stages, where you work out what to target and design a molecule worth chasing. Those stages cost a fortune, which is exactly why low-margin diseases got screened out before anyone ran a single trial.

This is a beat we have been running a lot lately, and it usually lands the same way. In June an almost-autonomous AI chemist improved a genuinely hard reaction in medicinal chemistry, and AI started helping physicians diagnose rare genetic disease in children. Real work, in a lab, years from a pharmacy. A policy brief from early June put the number on that gap: more than 200 AI-discovered drugs are already in clinical trials, and not one has been approved. So what got my attention here is who is writing the cheque. Anthropic is not selling a tool to a biotech and taking a fee, it is running the programs itself, at the end of the market where no payday is waiting. Everybody already knows a model can find a molecule. Nobody has shown one can make the arithmetic close on a disease that has never made a company any money, and that's the thing I'll be watching. Plenty of labs say AI will cure disease. I haven't seen one spend its own money where there isn't any, and I'd take that over another coding benchmark.

the-decoder July 2026 report Anthropic launches its own AI drug discovery programs and Claude Science bioscience product to tackle diseases Big Pharma considers unprofitable including neglected and rare conditions — the-decoder.com · July 2026

Why this matters: If you or someone in your family lives with a condition too rare to be worth a drug company's time, a frontier AI lab just pointed its own money straight at that gap instead of selling tools to somebody else who might. Be hopeful about that, and don't reorganise your life around it, because an announced program is not a cleared medicine, and the road from an AI-designed candidate to something a doctor can hand you runs through years of preclinical work, trials, and regulators. Action this week: Track the milestones rather than the announcements, which for anyone means one bookmark: watch for the first named target and the first preclinical candidate out of Claude Science, and keep an eye on the 200-plus AI-found drugs already sitting in trials, because when one of those finally gets approved, we'll know this actually works. If a rare disease touches your family, the practical move is to find the patient registry for it, since those registries are how AI-accelerated programs find the cohorts they need and how families hear about a trial early. Follow this one, and we will carry it as it moves.

the-decoder.com: Anthropic launches its own drug discovery programs to tackle diseases Big Pharma considers unprofitable (July 2026)
statnews.com: Anthropic wants to develop its own drugs (June 30, 2026)
theverge.com: Anthropic's Claude Science and AI drug development (July 2026)
endpoints.news: Anthropic debuts Claude Science, an AI product for bioscience (July 2026)

» What to watch this week

Whether Apple ever admits the AI server chip is in there. Apple is describing thirty billion dollars of silicon as Wi-Fi and Bluetooth parts, and the AI-datacenter piece is so far someone else's reporting, not Apple's statement. Broadcom's next earnings filing is where a custom-accelerator backlog of that size would have to surface, and that is the document to read rather than the press release.
Whether Meta's coding price survives contact with the bill. Watch for the tell that showed up the last time a headline rate went too low: the number on the pricing page stays put while a surcharge, a tighter rate limit, or a quietly smaller model appears behind it.
Which platform flips its Cloudflare AI-crawler default next. Patreon is the first big paid-creator platform to do it at the platform level, and the one to watch for is a publisher, because when a publisher does it, AI companies have to start paying for news or go without it.
Whether Anthropic names a first target, and whether any of the 200-plus AI-found drugs in trials gets an approval. The first approval is the event that would separate a genuine shift in discovery economics from a very well-funded press cycle, and it is the one number to keep an eye on all year.
Whether anyone else signs with Broadcom. OpenAI in June, Apple in July. A third name inside a quarter would say the custom-accelerator wave has stopped being a hyperscaler experiment and become the default way large AI buyers escape Nvidia's pricing.

Tomorrow's signal lands here.