Zuckerberg’s Contradiction, Samsung’s Chips, and the Adoption Numbers Behind Both
Mark Zuckerberg Admitted Meta’s AI Agents Stalled for Four Months. Minutes Later, His Own AI Chief Claimed an Unreleased Model Already Beats GPT-5.5. Meanwhile, Anthropic Opens Chip Talks With Samsung, and New Data Shows What Enterprise AI Adoption Actually Looks Like Underneath the Headlines.
At a July 2nd town hall, Mark Zuckerberg told Meta staff the company’s AI agents had stalled for four months — a rare, direct admission of internal struggle from a CEO not known for that kind of candour. Minutes later, according to accounts of the same meeting, Meta’s AI chief told the same room that an unreleased internal model, nicknamed Watermelon, already outperforms GPT-5.5. Meta shares fell nearly 5% on the reporting. Separately, Anthropic has reportedly opened preliminary chip discussions with Samsung, and new adoption data shows what enterprise AI use actually looks like two years into the boom. Ground Truth covers a contradiction, a hedge, and the numbers underneath both.
Admitting a four-month stall and claiming a category-beating unreleased model in the same meeting is not necessarily dishonest — both things can genuinely be true of different teams inside the same sprawling company. But it is a striking thing to say back to back, in front of the same room, and the market reacted to the tension between the two claims more than to either claim on its own.
Ground Truth · Episode 25 · July 14 2026At a July 2nd all-hands, Mark Zuckerberg reportedly told Meta employees the company’s AI agent efforts had stalled for roughly four months — an unusually direct acknowledgment for a CEO whose public posture on AI has otherwise stayed relentlessly bullish. Minutes later in the same meeting, according to multiple accounts, Meta’s AI chief told the room that an unreleased internal model, nicknamed Watermelon, already outperforms GPT-5.5. Meta shares fell nearly 5% following reporting on the town hall. Separately this week, Anthropic reportedly opened preliminary discussions with Samsung Electronics about a custom AI accelerator, and new adoption data gives a clearer read on where enterprise AI actually stands two years into the current boom. Welcome to Episode 25 of Ground Truth.
Four Months Stalled, Then a Model That Already Wins. Both, Somehow, Are the Message.
The stalled-agents admission is, on its face, a genuinely useful data point for this series’ running coverage of agentic AI reliability — Meta joining a growing list of companies discovering that shipping agents that work consistently in production is harder than a demo suggests. This series covered exactly that gap on Inside The Machine Day 18 and Day 31: the distance between an agent performing well on a curated benchmark and an agent performing reliably across the messy variety of real production workloads. Zuckerberg naming a four-month stall specifically, rather than deflecting with vague optimism, is a rare public data point in an industry where most companies say almost nothing concrete about their internal struggles.
The Watermelon claim, delivered minutes later in the same room, does not necessarily contradict the stall admission — a company can have one team’s agentic product stall while a separate research team’s unreleased model genuinely outperforms a competitor on specific benchmarks. But the juxtaposition, reported the way it was, reads as an attempt to immediately paper over an uncomfortable admission with a more flattering one, in front of the same audience, within minutes. Whether that was the intent or simply how a long town hall got compressed into a few reported lines is something only people in the room can say with confidence.
The market read it as a contradiction regardless of intent: Meta shares fell nearly 5% following the reporting, a reaction more consistent with investors hearing incoherence than investors hearing good news buried inside bad news. Whether Watermelon actually ships, and actually beats GPT-5.5 once it is independently benchmarked rather than internally claimed, is the test that will settle which version of this story was accurate.
Anthropic-Samsung: exploratory talks, 2nm process reportedly under discussion, aimed at reducing Nvidia dependency. Qualcomm-Tenstorrent: reported by The Information June 15, publicly denied by Tenstorrent’s CEO June 30, status unresolved. Meta: preparing to rent out its own AI compute rather than build custom silicon, a different diversification bet entirely. Three different labs, three different hardware strategies, one shared motive: reducing single-vendor dependency on Nvidia.
A Third Lab, a Third Approach to the Same Nvidia Problem
The Information reported on July 2nd that Anthropic has opened preliminary discussions with Samsung Electronics about manufacturing a custom AI accelerator, potentially built on Samsung’s 2nm process with advanced packaging. The talks are described as exploratory: Anthropic has reportedly hired specialised silicon engineers but is still defining chip specifications, power requirements, and how the resulting hardware would integrate into its existing server clusters. Nothing here is close to a signed agreement or a product roadmap.
What makes it worth tracking is the company it puts Anthropic alongside. This series covered Qualcomm’s reported (and subsequently denied) approach to Tenstorrent last episode, itself one attempt at reducing Nvidia dependency through acquisition. Samsung talks would be a different model entirely — a manufacturing partnership rather than an acquisition, closer in structure to how Tesla and Apple have pursued custom silicon in the past. Three different labs, three different structural approaches, one shared motive: nobody wants to be entirely dependent on a single chip vendor for the hardware their entire business runs on.
For Samsung, landing Anthropic as a high-profile foundry customer alongside Tesla would be a meaningful validation of its 2nm process and advanced packaging capability, at a moment when TSMC still holds a commanding lead in leading-edge foundry work. For Anthropic, even exploratory chip talks signal a company thinking past its current Amazon-anchored compute relationship toward genuine hardware diversification — the kind of long-horizon infrastructure bet that matters more in 2028 than it does this quarter.
An acquisition attempt, a denial, and now a manufacturing partnership under discussion are three different shapes of the same underlying anxiety: nobody building frontier AI wants to wake up one morning entirely at Nvidia’s mercy on pricing or supply. That anxiety is now expensive enough to reshape which chip companies get called first.Neal Lloyd · Ground Truth, Episode 25
From 33% to 71%, and Governance Is Doing the Sorting
McKinsey’s global surveys put the share of organisations using generative AI in at least one business function at roughly 33% in 2023, climbing to about 71% by 2024, and continuing upward into 2026. That is a genuinely fast adoption curve by the standard of prior enterprise technology waves — faster than cloud computing, faster than mobile-first redesigns. On its own, the number supports an unambiguous good-news narrative about AI diffusion.
The more useful read sits underneath the headline figure: adoption has shifted from isolated chatbot pilots toward AI embedded in core workflows — finance reconciliation, customer support triage, code review, marketing production — and the companies reporting real ROI are consistently the ones that redesigned processes around AI rather than bolting it onto existing workflows unchanged. That distinction between structural adoption and cosmetic adoption is exactly where this series’ Meta town hall story and the New York City Department of Education’s new bias-and-equity review requirement for its 1.1 million-student system both land: institutions that build governance and process discipline around AI report durable value; institutions that skip it report impressive demos that quietly fail once real usage begins.
That is the throughline across everything in this episode. Meta’s stalled agents, Anthropic’s careful hardware diversification, and the adoption data all point at the same maturing phase of this industry: the easy wins from simply having access to a capable model are mostly already banked. What separates the companies pulling ahead now is discipline in how they deploy it, not which model they happen to be using.
Seventy-one percent of companies using generative AI somewhere in the business is not the same story as seventy-one percent of companies using it well. The gap between those two numbers is where this year’s real winners and losers are actually being decided, and it has almost nothing to do with which model any of them chose.Neal Lloyd · Ground Truth, Episode 25
Ground Truth, Episode 25 · July 14 2026
Neal Lloyd covers the real-world impact of AI — money, power, geopolitics, and the stories behind the headlines. Ground Truth is his daily AI news and analysis series on emdexter.blogspot.com.
- Ep 01The Gold Rush
- Ep 02ChatGPT Knows Everything
- Ep 03Siri Is Now Google
- Ep 04America’s AI Law Is a Mess
- Ep 05Is AI Taking Your Job?
- Ep 06Microsoft vs Everyone
- Ep 07SpaceX Is Trading
- Ep 08The Government Pulled Fable 5
- Ep 09Trump and Bernie Want to Own AI
- Ep 10SpaceX Buys Cursor for $60B
- Ep 11The Fable 5 Truth
- Ep 12The Week That Changed Everything
- Ep 13Bots Now Outnumber Humans
- Ep 14Colossus: $80B Compute Landlord
- Ep 15ChatGPT Is Getting Ads
- Ep 16Alibaba Stole 28.8M Conversations
- Ep 17June 2026: The Month AI’s Rules Changed
- Ep 18The Credibility Gap
- Ep 19Sonnet 5, Fifty States, and the Return of Fable 5
- Ep 20Now Every Frontier Model Needs a Permission Slip
- Ep 21193 Countries, One Room, and a Compute Gap Nobody Can Vote Away
- Ep 22The Free Ride Ends, the Real Commission Meets, and Gemini Finally Shows Up
- Ep 23The Fable 5 Ban, Fully Told
- Ep 24The Denial, the Investigation, and the Cancelled Ceremony
- Ep 25Zuckerberg’s Contradiction, Samsung’s Chips, and the Adoption Numbers Behind BothYou are here
