Cc

Sony Kills the Disc, Anthropic's Fable 5 Is Back, and South Korea Bets $576B That Chips Are Everything

Rough week to be a physical-media hoarder. Sony dropped a bombshell on July 1 — it's pulling the plug on PlayStation game discs starting January 2028. That's not that far off. Meanwhile Anthropic's Fable 5 and Mythos 5 just got unbanned after an 18-day US export control standoff, and South Korea unveiled a jaw-dropping $576 billion semiconductor strategy. Plus GM locked down a direct chip deal with Micron, and there's a new AI inference breakthrough that might actually make your chatbots faster without wrecking your wallet. Let's dig in.

🎮 Sony Finally Draws the Line — Discs Are Done by 2028

Sony made it official on July 1: starting January 2028, no more physical game discs for new PlayStation titles. That means PS6 is almost certainly going digital-only, and PS5's disc-drive version might be the last of its kind. Honestly, this was coming. PC gaming already proved digital distribution works. But for the physical collectors, the folks who actually trade in games or buy used — this stings. Rockstar already skipped a disc for GTA VI, so the writing was on the wall.

From a practical standpoint: if you're sitting on a PS5 with a disc drive, you've got about 18 months of new disc releases left. After that? eBay prices for physical copies are gonna climb. One thing that bugs me though — Sony didn't say anything about backward compatibility for existing discs. If PS6 can't play my PS4/PS5 discs, that's a real middle finger to anyone with a shelf full of games.

🤖 Anthropic's Fable 5 Is Back Online — With Strings Attached

On June 30, the Trump administration quietly lifted export controls on Anthropic's Claude Fable 5 and Mythos 5 models. These were banned on June 12 over national security fears — the idea that state actors could weaponize frontier AI. The ban effectively cut off global users. Now Anthropic's slowly restoring access with a 50% weekly usage cap through July 7 and new safety classifiers in place.

To be fair, the ban itself was kind of chaotic. The US Commerce Department dropped it without much public explanation, then reversed course after 18 days. That's not exactly reassuring if you're a developer building on top of these models. Anthropic's now operating under formal government oversight, which sounds like a polite way of saying "we're being watched." For regular users? Fable 5 is genuinely impressive — I've been testing it on coding tasks and it's noticeably sharper than Claude 4 on long-context prompts. Whether that's worth the regulatory drama is another question.

🧠 South Korea's $576B Chip Plan — Samsung and SK Hynix Take Center Stage

South Korea dropped what might be the biggest single-country semiconductor investment package in history: $576 billion focused on AI chips, memory manufacturing, and regional industrial hubs. Samsung and SK Hynix are the main drivers. This isn't just about keeping up with TSMC — it's about owning the entire AI memory pipeline. HBM (high-bandwidth memory) is the bottleneck for every AI data center right now, and SK Hynix and Samsung make basically all of it.

Quick add-on note: this kind of spending is why DRAM and NAND prices are climbing. If you're planning a PC build this year, buy your RAM and SSD sooner rather than later. The Longsys earnings report confirms it — China's biggest memory module maker just reported a 600-fold profit jump. That's not a typo. Six hundred times. Tight wafer supply plus AI-driven demand equals expensive memory across the board.

🚗 GM Goes Direct to Micron for Automotive Chips

GM signed a long-term agreement with Micron to secure memory and storage chips directly — bypassing the spot market that's been running hot since the pandemic. This is a smart move for two reasons. One: cars are basically computers on wheels now, and every infotainment system, ADAS camera, and EV battery controller needs chips. Two: GM got burned during the 2021-2022 shortage and clearly doesn't want a repeat. The partnership also includes joint innovation on future automotive tech, which could mean custom memory solutions down the road.

For consumers? This probably won't make cars cheaper — but it might prevent the nightmare scenario where you can't get a new Chevy because some random sensor chip is backordered 14 months.

⚡ STAR-KV: Making AI Inference Suck Less

Dnotitia, a Korean AI infrastructure startup, just published STAR-KV — a low-rank KV cache compression technique that was selected as a spotlight paper at ICML 2026 (that's about 2.2% of submissions, for context). The headline number: up to 20x compression on the KV cache with only minimal quality loss. In practice, that means attention computation speeds up by 6.9x and overall generation throughput improves by 3.1x. What does that mean for someone using ChatGPT or Claude? Faster responses, longer context windows without hitting a memory wall, and potentially cheaper API costs since providers need fewer GPUs per query.

The paper and code are on arXiv and GitHub if you're into the technical weeds. But the takeaway for the rest of us: AI inference is getting real efficiency improvements, not just bigger models. That's the kind of progress that actually matters for day-to-day users.

🎬 Quick Hits

Netflix is using an AI-generated Gene Wilder voice (via ElevenLabs, with family consent) for its upcoming Wonka's The Golden Ticket reality show. Creepy? A little. Legally cleared? Apparently yes. Bending Spoons — the Italian company that bought AOL, Evernote, and Vimeo — is going public at a potential $19 billion valuation. And TV Time, the popular tracking app, is shutting down July 15 as its parent company pivots to enterprise AI. Feels like every consumer app with half a user base is chasing AI dollars these days.

Alright, where does that leave us?

If you're building a PC right now, pull the trigger on storage and RAM — prices are only going one direction through the end of the year. For gamers, the PS5 disc version is still a safe buy if you want to play physical through 2028, but don't expect the next console to have a drive. And for anyone using AI tools day-to-day, keep an eye on inference efficiency improvements — they'll quietly make your experience better without you noticing, which is honestly the best kind of tech upgrade.

What's your take on Sony going all-digital? Is losing physical discs a real loss or just nostalgia talking? Drop your thoughts below.

Netflix Cloned Gene Wilder's Voice, Sony Kills Game Discs, and South Korea Drops $576B on Chips

Saturday was slow but the news machine doesn't stop. Couple things worth unpacking — especially if you're sitting on a PS5 or wondering where your streaming dollars actually go.

Let's start with Netflix because this one's a bit wild. They're launching a Willy Wonka reality show called The Golden Ticket (September 23rd, mark it or don't), and the trailer voiceover? That's an AI-generated Gene Wilder. Netflix worked with ElevenLabs and got the family's blessing — they've done the same thing before with Michael Caine and Stan Lee voices. On one hand, the sets look real, no Glasgow Willy Wonka disaster this time. On the other, hearing a dead actor's voice read promo lines for a reality competition show feels... off. Not illegal, not unethical per se since the family signed off, but there's an ick factor that's hard to shake. The tech works, but maybe ask should we before can we next time.

Meanwhile over at Anthropic, the US government quietly lifted export restrictions on their Fable and Mythos models. These were suspended back in June over hacking concerns, and now they're back online as of July 1st. No fanfare, no press tour — just a quiet un-ban. If you're running any workflow on Claude, you might notice Fable-grade models becoming available again in certain regions. Good for Anthropic's bottom line, and frankly good for anyone who felt the rug got pulled too aggressively.

South Korea dropped what might be the biggest semiconductor announcement of the year — a $576 billion national tech strategy fronted by Samsung and SK Hynix. That's not a typo. Half a trillion dollars aimed at keeping Korea ahead in the memory chip race, with AI infrastructure as the main customer. DRAM and HBM prices have been climbing all year thanks to the AI boom hoovering up every wafer in sight, and this investment signals that Korea intends to stay the dominant supplier. If you've been watching RAM prices for a build, don't expect relief any time soon.

Speaking of chips and legal trouble — Super Micro's Taiwan unit has four employees under investigation for allegedly smuggling AI servers equipped with Nvidia chips. Document forgery, breach of trust, the whole deal. This is the kind of story that reminds you how tightly controlled advanced semiconductor exports really are. The US has been leaning hard on Taiwan to enforce chip export rules, and this case shows they're not messing around.

Sony confirmed what many of us saw coming: physical PlayStation game discs are done by early 2028. All-new releases will be digital-only. And here's the kicker — people dug into Sony's Terms of Service and found that PlayStation accounts inactive for 3 years in the EU can be deleted. That means your entire digital library, poof. So Sony wants everyone to go all-digital, but also reserves the right to wipe your account if you take a break. That's not FUD — it's literally in the ToS. If you're a physical media holdout, 2028 is your deadline. After that, every "purchase" is really just a rental.

There's also a quieter but relevant thread: Apple's Hide My Email feature has a bug that's apparently been exposing real email addresses instead of the masked ones. A researcher found it, Apple hasn't commented yet. For anyone relying on iCloud+ privacy features, worth keeping an eye on.

And on the business side, Bending Spoons — the Italian company that bought AOL, Evernote, and Vimeo like some kind of internet antique collector — is going public at a potential $19 billion valuation. That's either impressive or terrifying depending on how you feel about zombie brands being revived with AI glue.

Tesla is capping internal AI tool spending across the board — except for Grok, which gets an exemption. Funny how that works. Meanwhile employees reportedly prefer Claude over Grok internally, which is awkward when your CEO owns the competing chatbot.

For builders and buyers: if you're planning a PC build this summer, pull the trigger sooner than later. DRAM prices aren't coming down with Korea's $576B plan — that's long-term capacity, not this quarter. If you're sitting on a PS5 library, maybe log into your PSN account this week just to reset that inactivity clock. And if you're shopping for a used GPU, Super Micro's legal mess could ripple into server GPU availability in the coming months.

What's bothering you more — Sony deleting your digital library after three years, or Netflix resurrecting dead actors with AI voice cloning? Drop a comment, I'm genuinely curious where the line is for most people.

Let me be honest — this week felt like someone fast-forwarded the tech calendar. South Korea announced a jaw-dropping $576 billion semiconductor strategy, Sony confirmed it’s killing physical PlayStation discs by 2028, and Anthropic quietly dropped a new tool that lets scientists build AI research pipelines without needing a CS degree. Also, Micron signed a chip supply deal with GM, Dnotitia’s STAR-KV compresses AI memory by 20x, and Netflix is using an AI-generated Gene Wilder voice for a Willy Wonka reality show.

The $576 Billion Question South Korea just unveiled what might be the most ambitious semiconductor investment plan in history — $576 billion, led by Samsung and SK Hynix. The government wants to build new industrial hubs, lock down memory chip dominance, and push into foundry services. From my perspective, this is both impressive and slightly desperate. SK Hynix is crushing it with HBM, but Samsung’s foundry business has been struggling to match TSMC’s yields. Throwing money alone doesn’t fix process node issues — Intel spent billions on 7nm and still fell behind. The real question nobody’s asking: where will all the engineers come from? Korea’s population is shrinking, and chip talent doesn’t grow overnight.

Anthropic’s Claude Science Anthropic launched Claude Science — an AI workbench for researchers that connects Claude to data analysis tools, Python execution, and database queries. Think Claude + Jupyter Notebook built into the chat. Quick add-on note: OpenAI had Code Interpreter, Google has Gemini code execution. What’s different here is the target — scientists, not devs. To be fair, I’ve played with similar tools, and the gap between demo and daily use is still wide. Real research workflows need reproducibility and version control. Claude Science is a solid start, but don’t ditch your RStudio setup just yet.

STAR-KV: The Research That Actually Matters Dnotitia released STAR-KV — compresses KV cache by up to 20x while keeping inference quality. Accepted as a Spotlight at ICML 2026 (top 2% of submissions). KV cache is the silent bottleneck making long-context AI expensive. STAR-KV speeds up attention computation by 6.9x and generation throughput by 3.1x. This means cheaper API calls, faster responses, and eventually models that hold hour-long conversations without burning GPU memory.

Samsung Z Fold 7: One Year Later A TechRadar writer who’s been using the Fold 7 for a full year shared honest thoughts. The crease is barely visible, battery lasts a full day, multitasking is genuinely useful. But annoyances crept in: the outer screen is still too narrow for comfortable typing, cameras don’t match the S Ultra line, and micro-scratches appeared around the hinge after months of use. Interestingly, he said S Pen support isn’t a priority for the Fold 8 — most people just don’t use it on a foldable. If you’re thinking about foldables, the Fold 7 is the best option right now. But waiting 6 months for the Fold 8 might be worth it.

Sony Going All-Digital Sony confirmed starting in 2028, new PlayStation games won’t get physical disc releases. I have mixed feelings. Digital is convenient — no swapping discs, no scratched games. But this kills the used game market, removes resale ability, and ties your library to Sony’s ecosystem. Get your physical copies while you can.

Quick Hits - Micron and GM signed a long-term chip supply agreement for automotive memory — direct response to the 2021-2023 chip shortage. - Super Micro’s Taiwan unit: four employees detained over allegedly exporting AI servers with Nvidia chips in violation of controls. - Bending Spoons (AOL, Evernote, Vimeo) is going public at a $19 billion valuation. - Netflix using ElevenLabs AI to recreate Gene Wilder’s voice for a Wonka reality show, with family consent. Ethical lines are blurry.

If you’re building a PC right now, keep watching RAM and SSD prices — the AI memory crunch is real, and Samsung/SK Hynix’s huge investments won’t translate to cheaper consumer products overnight. Console gamers: grab physical copies of games you care about while you still can.

What’s your take on Sony going all-digital? Dealbreaker or no big deal?

If you blinked through the Fourth of July weekend, you missed a lot. The AI industry doesn't take holidays. Between Anthropic launching a new model at aggressive pricing, OpenAI floating a government equity stake that sounds like sci-fi policy, and the chip industry quietly pushing through another round of price hikes, there's plenty to unpack. Let's get into it.

The AI Price War Is Getting Real

Anthropic dropped Claude Sonnet 5 on June 30, and they're pricing it to move. At $2 per million input tokens and $10 per million output tokens through August 31, it undercuts their own Opus 4.8 ($5/$25) by a mile. My guess? They're trying to grab market share before OpenAI's next move. And OpenAI is definitely planning something — rumors of a price cut on GPT-series models have been circling since mid-June, and the Wall Street Journal reported OpenAI is weighing "steep reductions."

The subtext here is interesting. Both companies are burning cash on training runs, infrastructure, and now custom silicon. OpenAI just announced Jalapeño, their own inference chip built with Broadcom, and now Anthropic is in early talks with Samsung to manufacture a custom AI accelerator using Samsung's 2nm process. Two frontier labs, both deciding they can't rely on Nvidia alone for the long haul. Honestly, from a buyer's perspective, more competition in the chip supply chain is good news — it should eventually drive down compute costs for everyone running inference workloads.

But here's the catch. Palantir's Alex Karp went on CNBC on July 1 and blasted the whole token-based pricing model, saying enterprises and governments need to own their data, not pay per query. He's not wrong. If you're building an AI-powered product and your margins depend on someone else's token pricing, you're in a fragile position. The open-weight model camp (Llama, Mistral, DeepSeek, Qwen) is looking more attractive by the day for anyone who wants predictable costs.

Anthropic also launched Claude Science — a research workbench using multi-agent orchestration across 60+ tools, available to all paid subscribers. No enterprise vetting required. That's a solid move for researchers who want structured AI assistance without navigating procurement red tape.

Microsoft, meanwhile, is overhauling Copilot again. Internal memos say they're merging consumer and enterprise apps into one, adding "AutoPilot" agents that handle scheduling, email summaries, and other background tasks. They also stripped out Copilot Podcasts and Copilot Labs — features that were clearly not working. The executive in charge said the app needs to "earn the right to exist." Refreshingly honest, honestly. Microsoft also launched a new company that will embed 6,000 AI engineers directly inside enterprise clients to build AI into workflows. A chatbot alone, they've realized, isn't delivering enough measurable value.

Mark Zuckerberg told employees Meta's AI agent progress is slower than expected. That's a notable admission from someone who's been spending heavily on AI infrastructure. The gap between demo-ready and production-ready is real, and even Meta is feeling it.

The Chip World: Price Hikes, Memory Wars, and AMD's Big Bet

Over 20 global semiconductor companies kicked off a second round of price hikes in July. Power chip increases are running 15% to 25%, driven by AI demand and rising manufacturing costs. TSMC's 2nm wafers carry roughly a 50% premium over 3nm. Analog Devices raised prices 10-30%. OSAT packaging surcharges are up 8-20%. This is the broadest simultaneous price increase cycle in the chip industry in over a decade, and it's hitting everything from smartphone components to automotive chips.

If you're building a new PC or server, this is the worst time to be buying silicon. Memory is especially tight — HBM4 is basically pre-sold to hyperscalers, and the AI memory shortage that CNBC reported on in January is still unresolved. DDR5 prices are creeping up again. NAND isn't far behind. My advice: if you've been eyeing a 64GB or 128GB RAM upgrade, don't procrastinate. Lock in prices now rather than waiting until fall, because the second wave of chip price hikes hasn't fully hit retail yet.

AMD's Advancing AI 2026 event is scheduled for July 23, and the stakes are high. Cantor Fitzgerald just raised AMD's price target to $700 — the highest on Wall Street — citing AMD's momentum. The Helios platform, built around the Instinct MI455X GPU with 432 GB of HBM4, is going head-to-head with Nvidia's Vera Rubin NVL72. AMD leads in memory capacity per rack (31 TB vs 20.7 TB), which matters for large-model inference. But Nvidia still leads in training throughput (2.5 exaFLOPS vs 1.4 exaFLOPS at FP8) and has a more mature interconnect with NVLink 6.

The honest take? AMD's hardware is credible now. The software stack (ROCm) is still playing catch-up, but the gap is narrowing. If you're a data center operator, having a second viable supplier is a relief. For individual builders and gamers, the upcoming AMD and Nvidia consumer GPU lineups should benefit from the MI400 and Vera Rubin architectural work, but that's a 2027 story at best. For office users and general productivity, there's nothing urgent — current hardware is more than adequate for everything except heavy AI workloads.

Infrastructure Reality Check: Power, Politics, and Permits

Google's 11th annual Environmental Report, released June 30, showed electricity consumption rose 37% in 2025 — the largest single-year increase in the company's history. Total electricity use is up more than 250% since 2019. To their credit, they still managed to cut operational emissions 2% by matching 100% of electricity with renewable purchases. But water consumption also jumped 34%. The scale of AI's energy footprint is becoming impossible to ignore.

Blackstone-owned QTS killed its planned Digital Gateway data center project in Prince William County, Virginia, after years of local opposition, legal fights, and concerns over land use and energy demand. Data centers need power, water, and land — and communities are pushing back. The bottleneck for AI might not be chips or models. It might be public permission to build.

Nvidia is rolling out a "compute now, pay later" model for AI startups, letting cloud providers access GPUs through revenue-sharing structures instead of upfront payments. It's a smart move — gets more Nvidia silicon into the hands of startups that can't afford $100K+ GPU clusters upfront, and locks them into the Nvidia ecosystem early.

Quick Hits

- Unitree Robotics got approval for a $619M Shanghai IPO on the STAR Market. China wants physical AI out of the lab and into factories. - Google lost its final appeal against the €4.1B EU Android antitrust fine. Europe's top court upheld the ruling, strengthening the EU's position as Big Tech's toughest regulator. - Anthropic rolled back a Claude Code feature after scrutiny over time zone tracking linked to Chinese tech companies. The tension between security controls and developer trust isn't going away. - Apple is reportedly lobbying Washington for approval to buy cheaper memory from Chinese DRAM maker CXMT, according to industry sources. Unverified industry rumor, for reference only.

Alright, that's the weekend roundup. What's your take — are you buying the AMD Helios hype, or sticking with Nvidia's proven track record? And if you're running local AI models, how are you dealing with the memory crunch? Drop a comment, I'm curious what everyone's actually running in their rigs right now.

☕ Tuesday Tech Roundup — June 30

🗞️ Google dropped the paywall on Gemini's personalized AI image gen for US users — now free-tier users can generate images based on their Google data without typing detailed prompts. Opt-in, toggle-able. This is Google chasing ChatGPT and Claude on personalization, and 750M MAUs is the number they're protecting.

🤖 India is quietly pivoting to Chinese open-source AI models after US tightened restrictions on OpenAI and Anthropic exports. Pragmatic move, not political — when your AI project can't afford supply chain risk, you go with what works.

🧪 Researchers at Aalto University used machine learning to discover two new superconductors (YRu3B2 and LuRu3B2). The ML-first approach could scale to billions of candidates. Room-temp superconductor by 2033? Ambitious, but this pipeline is real.

🎮 Sony's next-gen PlayStation is going "beyond the living room" — corpospeak for a proper handheld or portable hybrid. PS5 Pro is already $900 and Xbox Series X is heading toward $750, so buckle up on pricing.

⚖️ SCOTUS ruled 6-3 that geofence warrants require a proper warrant with probable cause. Major privacy win. Police can't just draw a circle on a map and demand Google hand over everyone's location anymore.

📉 Comcast is spinning off all of NBCUniversal — NBC, Peacock, Universal Studios, theme parks — into its own public company. Media conglomerate era is over.

Apple raised prices 15–30% on Macs, iPads, and Vision Pro this week — blame the AI-driven memory crunch that's eating every DRAM and NAND chip in sight. Meanwhile, Google reportedly had to cap Meta's Gemini AI usage because it ran out of compute capacity, and Ford quietly rehired 350 veteran engineers after AI quality systems fell short. The RAM shortage is real, AI infrastructure is stretched thin, and even the biggest tech companies are scrambling — Apple is reportedly lobbying Washington for approval to buy cheaper memory from Chinese DRAM maker CXMT.