Signal Briefing: June 16, 2026

Tennessee Counties and Nashville Move to Ban New Data Centers

Multiple rural Tennessee counties have passed temporary data center moratoriums, and Nashville passed a near-unanimous ban on first reading, according to Tom’s Hardware. Three additional jurisdictions are set to vote on similar bills. The bans are framed as pauses to evaluate infrastructure strain before approvals resume.

Why this matters. Community opposition is now moving faster than permitting pipelines in some markets, creating a new category of siting risk distinct from grid connection queues or land costs. If Nashville’s moratorium survives later readings, it signals that even mid-tier metros — not just rural counties — are willing to slow hyperscaler and colocation expansion while local infrastructure catches up.

Confidence: high — Tom’s Hardware reporting on passed legislation; Nashville vote on the record.

---

TSMC Confirms CoWoS Stays Central as Panel Packaging Remains Years Behind

TSMC’s Kevin Zhang told attendees that panel-level packaging will not displace CoWoS for the largest AI processors in the near term, with wafer-level technology described as “considerably more advanced,” per Tom’s Hardware. TSMC is exploring its CoPoS panel technology but says wafer-level can already scale to 58 dies in a single package.

Why this matters. CoWoS capacity has been the single most cited bottleneck in AI accelerator supply for the past two years, constraining NVIDIA shipment ramps. TSMC’s statement sets expectations that any panel-based relief valve — which could meaningfully expand packaging capacity — remains a medium-term bet, not a near-term fix. Buyers planning multi-year AI cluster buildouts should not model panel packaging as a 2026–2027 supply unlock.

Confidence: high — primary TSMC executive statement at a public event, covered by Tom’s Hardware.

---

Oracle’s Project Jupiter Draws Water-Use Fire in New Mexico Desert

Oracle’s large-scale AI data center buildout in New Mexico, dubbed Project Jupiter, is drawing scrutiny from local officials and residents over water consumption in a region already under stress, Tom’s Hardware reports. Oracle characterizes the roughly 11 million gallon one-time fill as “negligible,” but local water-scarcity context makes that framing contentious.

Why this matters. Water is becoming a co-equal siting constraint alongside power for large AI campuses, particularly in the arid Southwest where evaporative cooling demand is highest. This parallels resistance patterns seen at hyperscaler sites in Arizona and the Pacific Northwest — and the Tennessee moratoriums above — suggesting a broadening coalition of local governments treating infrastructure density as a land-use and resource question, not just an economic opportunity.

Confidence: high — Tom’s Hardware reporting on disclosed project details; Oracle statement on record.

---

AMD Acquires MEXT to Bring Flash-as-DRAM Tiering to Data Centers

AMD has taken over memory tiering startup MEXT, gaining its Predictive Memory Engine technology that allows flash storage to appear as DRAM to applications, per Tom’s Hardware. AMD framed the acquisition as addressing “growing memory constraints” in data center deployments, particularly for large-model workloads.

Why this matters. HBM scarcity and cost remain hard limits on how much memory AI accelerators can carry economically, and DRAM bandwidth is a persistent bottleneck for large-context inference. A credible flash-tiering layer that is transparent to software would let AMD EPYC-based inference platforms effectively expand addressable memory without waiting on HBM supply chain expansion — a meaningful competitive lever against NVIDIA’s hardware memory advantages if the latency penalty can be kept within inference SLA tolerances.

Confidence: high — acquisition reported by Tom’s Hardware; AMD public statement cited.

---

APAC Regulators Shift to Grid-Support Obligations as Power Access Condition

A Wood Mackenzie report finds that regulators across Asia-Pacific are moving away from traditional utility interconnection deals toward frameworks that require data centers to provide active grid support services as a condition of power access, Data Center Dynamics reports. The shift reflects grid operators treating large AI loads as dispatchable assets rather than passive consumers.

Why this matters. This is the leading edge of a structural change in how power is priced and allocated to AI infrastructure globally. If grid-support obligations become a standard condition — as they appear to be moving in APAC — hyperscalers and colocation operators will need to engineer demand flexibility and storage into campuses that were previously designed for maximum uptime under stable utility supply. It adds capex and operational complexity but may also represent the only realistic path to securing gigawatt-scale power in constrained markets.

Confidence: medium — based on a single Wood Mackenzie report; specific country-level regulatory details not independently confirmed in this feed.