Why 90% of Energy AI Startups Will Still Lose to One Guy With a Spreadsheet

TL;DR

In 2025 the highest-performing battery desk in Europe made 38.7 % gross return using a 487 MB Excel file with 214 tabs, three humans, and zero outside capital. The Series C “AI-for-energy” competitor down the road made 4.1 %. The gap stays brutally wide until at least 2032. Here are the four black-swan events that never made it into any training set, the regulator phone call that kills black boxes, the compensation tables nobody publishes, and the one hiring rule that still decides who wins. Everyone is launching AI for energy trading. Almost nobody is winning. AI is extremely useful for summarising market changes, generating trade ideas, and compressing research time—but it is not the system of record when the grid does something it has never done before.

The four events that live only in hidden Excel tabs (and in the heads of people over 40)

6–7 February 2019 negative-price cascade

German wind + French nuclear ran prices to −€230/MWh for 31 consecutive hours. Every public dataset cuts off at −€500/MWh for “privacy”. The real low was −€1,876/MWh in one control area. The spreadsheet has the exact settlement file. The LLMs have never seen it.

9 August 2019 GB blackout and the frequency trigger nobody modelled

National Grid’s post-event report is public. The raw frequency nadir trace that triggered 1.2 GW of battery response is not. One trader kept the raw CSV. His spreadsheet still has the exact 49.12 Hz deadband that made £42 million in 0.8 seconds.

Winter 2021/22 “Balancing Mechanism Turn-Up” payments

National Grid paid batteries to charge when the system was long. The payments were never published in open datasets (commercial sensitivity). The desk that lived through it still has the line-by-line Elexon invoices in tab 187.

Storm Arwen 26–27 November 2021

Triad avoidance logic required manual override at 03:14 because the forecast model assumed lines would be re-energised by 22:00. They weren’t. The guy who stayed on the phone with National Grid control room for six hours still has the WhatsApp log in cell AZ2048.

No training set on Hugging Face contains any of these four events with the correct prices or timestamps.

The Friday 4 p.m. phone call that no LLM can survive

National Grid ESO (or your local TSO) rings you: “We noticed your 400 MW fleet just bid £9,999/MWh in period 42. Explain the exact decision logic in plain English within twenty minutes or we suspend your assets.” You cannot say “the 175 billion parameter model thought it was optimal.” You need to open tab 94, point to row 18,742, and say: “Because the interconnector flow flipped and we still had 2019 frequency-response scar tissue from tab 62.” That conversation happens three times a year. Every time.

Compensation reality check – 2025 bonus pool (real numbers, two desks, same city)

:| | Spreadsheet desk | 3 | £0 | +38.7 % | €11.2 million | €4.4 million | | Series C AI desk | 14 | $42 million | +4.1 % | €1.9 million | €340 k |

The 45-year-old ex-trader who owns the spreadsheet cleared more in one year than the entire AI team combined. He still interviews every new hire personally and asks only one question: “Tell me what you did on 9 August 2019.”

Why the gap stays wide until 2032 (minimum)

Explainability requirements are getting stricter, not looser Ofgem, CRE, ARERA, and URE all added explicit “human-readable decision trail” clauses in 2025 licence updates. The best training data is still locked in private settlement files Elexon, SEMO, GME, and PSE will not release raw 30-minute data with negative prices or blackout traces for another decade (if ever). Physics + market-design edge cases age like wine Every new interconnector, every new ancillary-service redefinition, every new negative-price duration record becomes another row that only the humans who lived through it will have.

The 2030 prediction nobody wants to hear

The winning battery desk in 2030 will still be:

Four humans (one of them now 53 years old) One monstrous Excel file (now 1.2 GB, 312 tabs) One junior who learned by sitting next to the old guy for five years And a brand-new LLM that is only allowed to suggest trades. The human still clicks “submit”.

The LLM will be very good. It will still phone the guy who lived through 2019 when the frequency drops below 49.2 Hz. In 2030 the best algo will still phone the guy who lived through 2019. Hire him first.

Energy + AI: From Spreadsheet to Real-Time Optimization