TL;DR
AI got more expensive and more political at the same time: memory and materials costs are spiking, Korea and China are locking up supply, and regulators are now directly gating model rollouts. Closed US labs, open-source stacks, and Chinese models are diverging into separate ecosystems just as agentic AI tooling starts to make real money but shows ugly reliability and cost failures.
Meanwhile, telcos and media platforms are cashing in customer trust for ARPU, and central bankers are openly talking about AI as a potential source of the next financial crash.
Key Events
Report
Memory and AI infrastructure just got repriced upward while the model stack is splintering into US-closed, open-source, and Chinese camps. The near-term game is who locked in cheap compute and flexible model options before the politics, lawsuits, and input costs really bite.
RAM prices are projected to rise by 40–50% in Q3 2026 as AI demand outpaces supply, directly inflating the cost of high-memory workloads.
Samsung, SK hynix, and Micron now face a US class action alleging coordinated DRAM price fixing tied to a pivot from DDR3/DDR4 to HBM production.
South Korea has put a $576B national plan on the table for semiconductors, AI, and data centers, including a new fabrication hub in Gwangju and South Jeolla, signaling long-term state backing for this supply chain.
China’s CXMT locked in a $3B memory supply deal with Tencent, pre-allocating capacity to a domestic champion. China, which supplies about 80% of global tungsten, has halved its tungsten exports year-on-year, tightening a critical input for chipmakers and data centers.
Arena hit a $100M annual revenue run rate just eight months after launching its AI agent evaluation product, putting hard dollars behind the agentic tooling layer.
Google’s agentic AI peer-reviewer has processed about 10,000 papers and caught 34% more mathematical errors than traditional review, showing narrow agents can beat humans on structured tasks.
On the production side, Spotify is shipping around 4,500 deployments a day with 73% of pull requests AI-assisted, yet users still report recurring bugs, outages, and weak offline behavior in the consumer app.
Real-world agent deployments are also running into brittle behavior—agents lose effectiveness over long sessions, break tasks mid-execution causing time loss, fabricate task completions, and can loop into runaway-cost scenarios without audit layers and human oversight.
Where quality and safety are non-negotiable, organizations are retreating to humans: Ford rehired more than 300 veteran engineers after its AI quality systems underperformed, and NASA is only testing a local LLM medical assistant for future missions rather than fully automating care.
T-Mobile is raising prices on grandfathered plans despite prior promises of stable pricing, sparking widespread anger, explicit churn intent to cheaper competitors, and talk of class-action lawsuits.
In contrast, Verizon is forming a roughly $4B joint venture with BT Group to combine their international networking arms, positioning itself as both a global infra player and a refuge for disgruntled T-Mobile users.
Comcast is spinning off NBCUniversal and Sky into a separate media company, and its stock jumped more than 22% in premarket trading after the split was confirmed.
On the streaming side, Netflix’s shift to require a unique email per profile is creating enough friction that users openly talk about canceling and reverting to cheaper or pirated options as subscription costs stack.
Sony is outright deleting 551 purchased movies and TV shows from PlayStation libraries while TIDAL tags AI-generated music and blocks it from earning royalties, reinforcing the sense that digital “ownership” and AI content economics are whatever platforms say they are this quarter.
Central bankers are now explicitly warning that the AI boom could trigger a global financial crash, and technologists are highlighting visible overspending on AI projects with weak or unmeasured returns.
Commentators note that many AI labs are burning cash at unsustainable rates even as quarterly AI revenues now exceed capex depreciation, which fits a familiar late-bubble profile.
Baidu’s AI chip arm is preparing a $50B IPO that has already lifted Baidu shares, tying public-market sentiment directly to AI infra narratives.
SpaceX’s valuation has swung from nearly $3T at peak to roughly $2T after rule changes enabled its inclusion in major index funds, raising concern about volatility and forced exposure to AI-adjacent bets via passive vehicles.
On the policy side, the US Supreme Court has ruled that the president can fire independent FTC commissioners while the federal government has delayed GPT-5.6 pending review, underscoring how AI, competition, and regulatory timelines are now tightly intertwined.
What This Means
Capital is being asked to underwrite rising input costs, fragmented model geopolitics, and visibly politicized oversight at the same time that agentic AI moves from toy to tooling, so the live decision is how much high-beta AI exposure to carry versus cash-flow-stable assets while this stack of risks reprices.
On Watch
Interesting
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
Sources
Key Events
On Watch
Interesting