RAMageddon Is Reshaping Tech (It’s Not Pretty)
AI Inside for Wednesday, June 30, 2026
This week on the AI Inside podcast, Jeff Jarvis and I tackled a packed show anchored by two massive stories: the global memory crisis now called RAMageddon that just sent Apple prices soaring, and the unprecedented week where both Anthropic and OpenAI released their most powerful models under government restrictions.
But first, a huge thank you to our patrons D J Rout and Fred Davisson for supporting us directly at patreon.com/aiinsideshow. You make this show possible.
RAMageddon
This story has been building for months, and this week it all came to a head in a few ways. Jefferies put out a research note warning that memory prices will surge 40 to 50 percent in Q3, another 30 to 40 percent in Q4, and no real relief until 2028.
For some context that was useful to me in indurstanding the scope of this: data centers now consume around 70 percent of all memory chips produced worldwide. A single NVIDIA B300 GPU requires 96 DRAM dies. A loaded DGX B300 system with eight GPUs has 768 dies in one server rack.
Samsung, SK Hynix, and Micron control over 95 percent of global DRAM production, and about 50 percent of total memory capacity is locked up in long-term contracts with big tech companies. That share could rise to 70 percent, meaning even less supply available for consumer products.
And it hits home when Apple is raising prices across a large swath of its product lineup. MacBook Air jumped $200. MacBook Pro 16-inch, up $300. The base iPad up $100. And the wee little Apple TV went from $129 to $199. This is the first time Apple has raised prices across this many product lines at once. Tim Cook even called it a “hundred-year flood” for memory and storage costs.
The thing is, Apple can do this. Apple has cushion, options, and 1.5 billion users locked into its ecosystem. The companies that can’t? That’s where things get ugly.
GoPro filed a regulatory warning saying there is “substantial doubt about the company’s ability to continue as a going concern.” Their memory costs shot up 80 to 115 percent at the end of Q1. Revenue declined 26 percent, and they’re laying off 23 percent of their workforce while evaluating a potential sale.
IDC is projecting the worldwide smartphone market will decline 13 percent in 2026, the largest drop ever, roughly 160 million fewer smartphones shipped year on year. The sub-100-dollar smartphone segment, around 171 million devices, becomes “effectively uneconomical.” I imagine the low-to-mid-tier smartphone world is about to see some attrition, and it might make a lasting impact on some of these companies.
The Week of Restricted Power
Something happened this week that I don’t think we’ve ever seen before. Two major US AI labs released their most powerful models in the same week, and both were released under government restrictions.
The Commerce Department lifted its block on Anthropic’s Claude Mythos 5, clearing it for about 100 US institutions, companies, and government agencies as part of Project Glasswing. The specific company names haven’t been made public. The consumer version, Fable 5, is still in limbo with no timeline. (NOTE: The Commerce Department gave the green light to the release of Fable 5 after the podcast recorded.)
Same week, OpenAI released GPT-5.6 in three versions: Sol (most powerful), Terra (balanced), and Luna (fast and cheap). But it’s also restricted, available as a limited preview to about 20 companies whose participation was approved by the government. The models are not available in ChatGPT during that preview. The partner vetting requirements are intense: continuous personnel screening, sovereign US cloud infrastructure, and real-time auditing hooks that log every prompt and output to a government-monitored data pool.
And then there’s the China angle. China’s Zhipu AI released GLM-5.2 under an open-weight license, free and running on consumer-grade hardware. The WSJ ran a headline saying “China Has Matched Anthropic in Cybersecurity, Resetting AI Race,” though Zvi Mowshowitz published a detailed rebuttal calling that headline “outright false.” He argues what makes Mythos special is not finding individual vulnerabilities but finding them autonomously, at scale, and stringing unrelated vulnerabilities together into full working exploits. GLM-5.2 cannot do that, but even Zvi acknowledges the trend matters.
And then China food delivery company Meituan released its own open-source model LongCat 2.0, which they say is the world’s first trillion-parameter model trained on a 50,000-chip local cluster, ie., no Nvidia. China continues to get much more self-sufficient in the process.
Google Caps Meta’s Gemini Use
The Financial Times reported that Google told Meta back in March that it could not supply all the Gemini capacity Meta wanted, and the restriction is still in place. What was Meta using Gemini for? Content moderation, removing harmful content and wiping out scams. Meta was using Google’s model because Gemini was better at these tasks than Meta’s own Llama open-source models.
Meta had to go to a competitor for AI models because its own weren’t good enough for safety tasks, and that competitor said sorry, we don’t have enough compute for you. Meta’s response has been to shift workloads to Muse Spark, a new internal model out of their Superintelligence Labs division.
Speed Round
Qualcomm’s Data Center Play — Qualcomm announced a server CPU called Dragonfly C1000, more than 250 custom Arm cores, claiming 2x performance per watt versus the competition. Mark Zuckerberg showed up in person to announce a “multigenerational agreement” for Qualcomm to supply CPUs for Meta’s next-gen servers starting in 2028.
Gemini 3.5 Flash Gets Computer Use — Google baked computer use directly into Gemini 3.5 Flash as a native built-in tool. Previously this required a separate standalone model. Now a single agent can see a screen, search the web, use Maps, and interact with interfaces without switching models. Computer use is becoming a standard capability.
Big Companies Launch AI Workforce Transition Effort — A new nonprofit called RAISE US launched with over 500 million secured, targeting 1 billion. It’s backed by Amazon, Anthropic, Microsoft, OpenAI Foundation, Bank of America, GM, IBM, Eli Lilly, and others. Nearly 50,000 job cuts have been announced in 2026 linked to AI, about 17 percent of all job cuts this year.
OpenAI and Broadcom Unveil Inference Chip — OpenAI announced its first custom silicon called Jalapeno, built on TSMC’s 3-nanometer process. Design to tape-out happened in 9 months, which they claim is the fastest ASIC cycle ever in high-performance semiconductors. It’s inference-only, not training, betting that as AI shifts to mass consumer products, the bottleneck becomes inference cost and speed. Deployment is expected at the end of 2026 at “gigawatt scale.” It’s notable that every major AI company now has custom silicon: Google has TPUs, Amazon has Trainium, Meta has MTIA, and now OpenAI has Jalapeno.
Executive Producers of AI Inside: DrDew, Jeffrey Marraccini, Radio Asheville 103.7, Dante St James, Bono De Rick, Jason Neiffer, Jason Brady, Anthony Downs, Mark Starcher, and Karsten Samaschke!
Thank you for watching, listening, and reading. See you next Wednesday on another episode of AI Inside.