
Research from @creativestrat analyst-grade intelligence on the companies reshaping the technology landscape. We analyze the full stack—from silicon to software.
Joined March 2025
- Tweets125
- Following8
- Followers1.1K
- Likes7
AI looks less like a new installation cycle than the deployment phase of the Information Age: the networks and devices already exist, so spending adds compute, memory, power and cooling on top. thediligencestack.com/p/welc…
Onsite power is not just an equipment problem. Reliability requires redundancy, firm fuel and longer customer commitments. Each raises cost, while an outage puts a far larger compute investment at risk. thediligencestack.com/p/expe…
Enterprise AI economics improve with scale because integration, data, security and staffing costs are front-loaded. Of scaled operators measuring full cost, 11 of 12 were at break-even or better. thediligencestack.com/p/the-…
A weaker PC market frees leading-edge wafers, but packaging and power still cap accelerator shipments. Server CPUs benefit more: they use less advanced silicon, need no HBM and avoid the largest packages. thediligencestack.com/p/titl…
120
Agents turn DPUs from a cloud tool into the AI data center’s control plane. As every agent becomes an untrusted virtual user, isolation, networking and security move onto an Arm-based server within the server. thediligencestack.com/p/exec…
126
Security is now the production gate for enterprise AI: 34% of recently reviewed projects were held mainly for security, yet only 16 of 49 held projects had funding approved to fix the gap. thediligencestack.com/p/the-…
123
Agent demand is not another web workload. By 2030 our Base case needs 30.1 GW of effective accelerator capacity against 27.3 GW available, because long histories make memory and power the bottlenecks. thediligencestack.com/p/we-n…
1
130
AI clusters depend on synchronized time. As links reach 200G and 400G per lane, clock error leaves less margin and can force links to retrain, turning timing into a utilization problem. thediligencestack.com/p/siti…
140
Chip-equipment demand is becoming supply-paced. Cleanrooms, energy and ASML output now limit installations, stretching the cycle across more years even if wafer demand holds. thediligencestack.com/p/chip…
AI infrastructure only pays when capacity is busy and monetized. In our central case, moving the paid share of the fleet from 42% to 50% is enough to cover the 2030 annual bill, making absorption the key variable. thediligencestack.com/p/can-…
Credo can optimize SerDes, DSP, photonics and firmware together, giving customers ways to trade power, reach and reliability. The thesis earns out when qualifications become repeat orders, gross profit and cash. thediligencestack.com/p/cred…
Agents expand CPU work beyond GPU hosts: tool environments run code and business services absorb every call. The opportunity exists only if useful work outruns efficiency gains and spare capacity. thediligencestack.com/p/agen…
154
NVIDIA’s roadmap, supply chain and co-design bench do not leave with one executive. Watch senior engineering departures or any softening of Rubin and the 2027 roadmap on Nov. 17; those would change the thesis. thediligencestack.com/p/nvid…
129
Hyperscalers need not move every workload between architectures to gain leverage. Qualifying custom silicon and alternate suppliers puts the next order in competition, with software deciding where work runs. thediligencestack.com/p/the-…
175
Memory cycles are waves, but bits are the tide. Even as realized DRAM pricing falls 39% from 2027 to 2030 in our Base case, bit shipments rise 89%, holding conventional DRAM revenue roughly flat. thediligencestack.com/p/for-…
164
Enterprise agents create less new data than expected; the bigger problem is making records reachable, governed, permission-mapped and current. Vendors controlling those gates can meter every action. thediligencestack.com/p/the-…
127
Memory suppliers share similar peak-cycle multiples, but implied downturns differ. SK hynix has the deepest discount despite the highest HBM mix—the opposite of what a firmer floor should imply. thediligencestack.com/p/bet-…
188
When power is capped, goodput becomes capacity. Meta reports Muse pretraining above 85% goodput; a ten-point improvement can equal roughly 50–75 MW of effective compute on a gigawatt-class site. thediligencestack.com/p/prom…
1
181
Agentic CPU racks sit outside the GPU scale-up domain, so they create a separate scale-out network. Add untrusted agents and DPUs become the isolation and control layer for that new CPU fabric. thediligencestack.com/p/exec…
The E/AI Index V2: Security Sets the Pace of Enterprise AI
Our survey of 200 security leaders finds security deciding which AI projects reach production, while many agents already act without basic controls
thediligencestack.com/p/the-…