ICELT Research

Research.

Independent market intelligence on AI compute, infrastructure economics, and emerging technology. Sourced, numbered, and built to inform decisions not to fill a newsletter.

Portfolio Governance · May 2026

The Funnel That Isn't.

Portfolio management promises selection. In most organizations the budget is spent before selection begins.

9 pages Capital allocation Budget sequencing Beyond Budgeting Public-sector evidence
  • The budget is fixed top down in spring, before the projects it funds are defined in autumn. The October sign-off ratifies a number decided in a room with no projects in it the funnel selects against a total that was already set.
  • McKinsey found year-to-year capital allocation correlates 0.92, with the median firm reallocating about 1% between units annually. Freeze the envelope before the contest of ideas and there is nothing left for a funnel to select.
  • The frameworks miss it. Stage-gate, PMO playbooks and the scaling frameworks can be adopted in full while the budget calendar stays exactly as it was. They govern execution hygiene, not the funding model.
  • The fix is old and proven. Handelsbanken has run without an annual budget since 1970 and beaten its peer group on return on equity in 46 of 47 years; the firms that reallocate capital most aggressively delivered ~30% higher shareholder returns over fifteen years.
0.92
Year-to-year capital allocation stickiness
1 in 6
IT projects run 200%+ over budget
30%
Higher TSR for aggressive reallocators, 15 yrs
Talent & Retention · May 2026

The Eighteen-Month Exit.

Dev hubs train juniors well and lose them anyway. The fix is not a bigger cheque.

10 pages Junior developer retention Career development Autonomy & tooling Mentoring evidence
  • Median tenure for workers aged 25–34 is just 2.7 years, about half that of workers aged 55–64. The exit window opens well before any resignation letter.
  • 41% of departures cite lack of career development ahead of pay (McKinsey, 12,000 respondents). A junior who cannot picture the next five years here will take a certain gain elsewhere over an uncertain future inside.
  • Developers spend only 11% of actual time coding against a 20% ideal. Micromanagement and ticket admin are a daily signal about how far the company trusts them and a daily reason to leave.
  • Replacing a leaver costs 0.5 to 2× annual salary. The counteroffer is the most expensive and least effective moment to act; 70% of internally promoted staff are still with the employer after three years, against 45% who stayed in the same role.
2.7 yrs
Median tenure, workers aged 25–34
41%
Quit for career development gap, ahead of pay
70%
Internally promoted, still there after 3 years
Market Intelligence · May 2026

Renting the Pickaxe.

The operator's side of GPU lending. Thin coupon, thick risk, and where the "low-risk" pitch breaks.

14 pages H100 / H200 / B200 Unit economics Risk decomposition 2026 outlook
  • A $1,500/month fixed H100 lease is structurally insolvent at mid-2026 spot prices of $1.49–$2.50/hr the lease was priced against a market that no longer exists.
  • Break-even utilisation on a 50/50 revenue-share in Texas is 27%. In Germany it rises to 46%. The gap between those two power costs ($0.06 vs $0.18/kWh) exceeds the combined colo and labour bill.
  • "Low risk" is technically honest on CapEx and materially misleading on operating risk utilisation, price compression, and repossession are all borne by the operator, not the hardware owner.
  • H100 rental pricing re-tightened 15–20% month-on-month into spring 2026. The expected oversupply did not arrive. This is the single most important data point for anyone underwriting a deal today.
$1.49
H100 spot floor / GPU-hour, mid-2026
46%
Break-even utilisation, 50/50, EU power
18mo
H100 active obsolescence lifecycle

Questions on the research or the methodology? Get in touch.