Actionable insights — Greg Abel on the Alphabet block and the AI power constraint

2:14 1. Price a block, not a share — the discount is the day-one edge

1. Distinguish the two ways a position gets built. Open-market accumulation pays the market price and is limited by daily volume; a negotiated block in an issuer's equity offering is a single transaction at a price you help set. Only the second is available to a buyer who can clear the whole line at once — and it is the one worth preparing for.
2. Recognise the moment the option appears: an inbound call, often outside market hours, in which neither the size nor the price is fixed. Treat that as a negotiation opening, not a term sheet. "Really, no terms or amount were set."
3. Answer with both variables yourself. Name the size you are willing to take and, separately, the discount to the prevailing market price you require for taking it — then let the issuer accept or refuse the pair.
4. Understand what you are being paid for. The discount compensates for immediacy and absorption — the issuer avoids weeks of market impact — so the correct size to bid is the largest one you would want anyway, because bidding small forfeits the leverage that produces the discount.
5. Separate the two decisions cleanly: do I want this business at all (a slow judgement made over months of open-market buying) and is this a good block (a same-day judgement about size and discount). The second only makes sense once the first is already settled.
6. Do not let the discount substitute for the thesis. A 6.5% concession is a single-quarter's worth of return on a name you would not otherwise own.

• Announced equity offerings, PIPEs and registered directs at large caps — the 8-K/prospectus discloses the block price against the prior close, which gives you the realised discount. A widening discount across deals is a tightening-liquidity signal; a narrowing one means anchor buyers are competing. Also watch which anchors keep getting the call — the callback list is itself the franchise.

2:42 2. The size-threshold rule — delegate the decision, escalate the magnitude

1. When authority is handed over, ask what the escalation trigger actually is. The useful distinction is magnitude, not subject matter: the decision maker owns every call in the category, and a defined size above which a second signature is sought anyway.
2. Test the rule against behaviour rather than the org chart. If the same person made many un-escalated decisions on the same name and escalated only once, the trigger is size — which is exactly what you want, because it means the delegation is real.
3. Note that the escalation call is a consultation, not an approval gate: "Warren and I discussed the size. We discussed the size of discount" — the recommendation still originated with the decision maker.
4. Apply this when assessing any succession: ask whether the successor has already been executing unsupervised in the category, and whether the predecessor's remaining involvement is triggered by a threshold or by veto over content. The first is a completed handover; the second is not.

• In any post-succession company: whether new large positions and acquisitions are announced with the successor's name on them, and whether the predecessor's public commentary shifts from directing to describing. A handover that is real shows up in 13F additions the predecessor did not initiate.

4:01 3. Use your own operating businesses as the demand channel — the conglomerate as a research edge

1. Before underwriting a technology story from outside, ask whether you already have an internal view of the demand. An owner of many operating businesses can observe adoption directly instead of inferring it from vendor revenue.
2. Score the technology on two questions your own operations can answer: are we using it, and what type of benefit is it delivering — cost, throughput, headcount, error rate. Adoption without a measurable benefit is not a demand signal.
3. Only then move to the security. The internal evidence establishes that spend is real and durable; picking the vendor is a separate judgement about competitive position.
4. Be honest about the limits of what the internal read tells you. It sizes the category; it does not rank the players — which is why Abel's stated reasons stop at "a significant player" and he concedes "there's a lot more to Google than what I just said."
5. The transferable form for an outside investor without a conglomerate: substitute your own firm's or industry's adoption, supplier conversations and hiring data for the internal channel — the point is to source demand evidence from operations rather than from sell-side forecasts.

• Operating-company commentary in conglomerate and insurer filings that quantifies internal AI use (claims handling, underwriting, logistics scheduling). Also watch for the reverse tell — a large diversified owner that is not adding AI exposure despite seeing its own usage is telling you the benefit is not showing up.

5:58 4. The utility's four-part test for accepting hyperscaler load

1. Do not model data-centre load as unconstrained demand a utility will always take. Regulated utilities apply a screen, and the screen determines how much of the announced pipeline converts. Reconstruct it before forecasting load growth.
2. Test 1 — ratepayer impact. The floor is "no impact to the rates of our other customers"; the standard Berkshire actually applies is higher — "there has to be a net benefit to our customers." Ask, for any utility you own, which of the two it is committing to, because that determines whether new load is accretive or merely neutral to the existing book.
3. Test 2 — water. "The communities have to understand the impact on water." Note that this constraint is improving as cooling technology changes — "much more manageable as they address that and use the technologies that are available to minimize water use" — so treat it as a diminishing rather than a fixed obstacle.
4. Test 3 — community consent. "The communities have to be open to having the data centre in their community… you have to be a welcomed member of the community." This is a project-underwriting variable, not public relations: it can kill a sited project after capital is committed.
5. Test 4 — pre-agreement with the regulator. The decisive procedural point: the principles are "really policy… we've discussed with our state, our governors and our regulators" and shared with every hyperscaler in advance. A utility that has pre-negotiated the terms faces far less rate-case risk than one improvising deal by deal.
6. Then score the customer, not just the utility: the hyperscaler chooses the site, so the utility can only "encourage them to seriously evaluate the reaction from the communities." Siting risk therefore sits with the tenant and shows up in timelines, not in the utility's returns.

• Large-load tariff filings and special-contract rate schedules at state PUCs (the mechanism by which "net benefit to other customers" is actually implemented — minimum-take provisions, exit fees, collateral). Also: county-level zoning rejections and moratoria as the leading indicator of Test 3, and utility disclosure of the ratio of requested to contracted data-centre load — the gap Abel names is the real forecast.

4:53 5. Reframe "energy is the constraint" — ask whether it is the electrons or the site

1. When a bottleneck is asserted, split it into capacity and deliverability, and ask an operator which one binds. These imply completely different trades: a capacity constraint favours generation and fuel; a deliverability constraint favours interconnection, transmission, transformers, engineering and construction, and lead-time-holding incumbents.
2. Take the answer from someone with a build record rather than a model. Abel's authority here is decades of infrastructure construction at Kiewit and then running the utility — his view of the constraint predates the AI cycle ("I've sort of always had a strong view that energy would be the constraint").
3. Test the framing against the operator's own words: "we can produce the energy" — capacity is not the problem — "it's… how long it would take to get the sites prepared and being in a position they could serve the data centres. And I continue to see that as a big constraint."
4. Convert the reframing into a measurable quantity: the gap between load requested and load servable. That spread, not installed generation, is the number to track.
5. Re-run the same split on any capacity narrative — refining vs crude, mine supply vs smelter capacity, chips vs packaging. The commonly named shortage is frequently the wrong link in the chain.

• Interconnection queue lengths and withdrawal rates at the major ISOs, transformer and switchgear lead times, and utility capex guidance split between generation and transmission/distribution. A shift in that split toward T&D is the constraint confirming itself.