Wednesday, December 31, 2025 · 3 newsletters

The Bottleneck Moved

agent-architecture · power-as-bottleneck · ai-as-commodity · year-end-reflection · payments-meets-agents · epistemic-anchoring

Roughly 150 emails across four publishing weeks, weighted toward year-end stocktaking. Volume was light by any measure (40, 32, 54, 23). No single news cycle dominated. What did emerge, across writers who do not normally talk to each other, was the cleanest architectural consensus of the year on where AI work actually sits in a team, plus a quieter macro thread on power, wages, and the gap between the rate-cut consensus and the data. The thin weeks rewarded the writers who picked a frame and committed.

The Month in One Sentence

This was the month the AI operator class stopped arguing about what the models can do and started arguing about where the human judgment sits in the loop, while the bond market and the grid quietly priced the physical constraint nobody on the software side wanted to talk about.

Arc: From Prompt Library to System Architecture

The month's structural arc on AI started with a prompt-band-aid critique and ended with a fully formed argument about org design. The shift inside four weeks is the cleanest tell of how fast operator consensus can re-classify a topic.

Week one was the architectural opening shot. Alex Furmansky at Magnetic Growth wrote the cleanest version: teams hit a problem, append ALL CAPS directives to the system prompt, repeat, and end up with a 20,000-token Frankenstein that contradicts itself. The fix is not a better prompt, it is better architecture. Nikunj Kothari at Balancing Act ran the CEO companion in "Bet the farm," attacking "transformation theater" and naming Tobi Lutke as the example of why the Chief AI Officer should be the CEO. Daniel Pupius at The General Partnership found the same split in engineering interviews: strong engineers used AI to scaffold boilerplate, weaker ones could not answer "what does this line do?" The convergence across three vantage points was already the signal of the month.

Week two was the commodity argument crystallizing. Yue Zhao at The Uncommon Executive named it for product managers: "As AI can now prototype, design, and perform data analysis, technical skills are increasingly a commodity. Your edge at work is the visibility, alignment, and agency you bring." SeattleDataGuy, hosting Mehdi Ouazza, ran the data-engineering version: "Coding matters less than ever before. Understanding how components work, their foundations, and their trade-offs, that's irreplaceable." Two writers, opposite ends of the table, the same conclusion. Pair that with Elena Verna on why freemium is not dead and Kyle Poyar at Growth Unhinged flagging the early AI churn wave, and the operator-level question for 2026 was already named.

Week three was Pupius writing the definitive essay. His GitHub integration triages issues with Claude and either produces a mini spec or a working PR; his role is no longer writing code but deciding what gets merged. When code generation gets cheap, the constraint shifts from "what can we build" to "what should we release," and the apparatus built around story points and sprints stops making sense. Addy Osmani opened the same week with the data point that ~90% of Claude Code is now written by Claude Code itself. Ethan Mollick at One Useful Thing extended his "Jagged Frontier" framing on permanent memory. The Graphite-Warp podcast episode, released the same day Graphite was acquired by Cursor, pinned the bottleneck on code review.

Week four collapsed the arc into infrastructure. Dwayne Gefferie's Payments Strategy Breakdown "Future Proof" argued 2025 was the year four parallel trends in payments matured at once, with the October Agentic Commerce Protocol launch as the inflection event. Alephic's Forward Deployed podcast on context engineering provided the operator-level primitives: reducing context, isolating context, offloading context to file systems. The line that stuck: "treat agents like humans works, because of dual-use tools and new coworker onboarding." The architectural argument that started with Furmansky's band-aid critique ended four weeks later with payments rails and protocol launches as the second half of the same story.

Arc: The Grid Is the New GPU

The month's sharpest non-consensus call ran across only three weeks but produced the cleanest physical-constraint frame of the year. Most of the AI capital story has been written from the software side. December was the month a small cluster of writers insisted the bottleneck is power, cooling, and the substrate beneath the chips.

Week one was the napkin math. Stonebridge Capital, "The TWh Gap: Why the Grid is the New GPU," argued the market has priced silicon to perfection and overlooked the physical bottleneck. Take 1.5 million H100-equivalent chips at 700W TDP, multiply by a 1.5 PUE for cooling overhead, run them 24/7, and you get roughly 13.8 TWh of annualized demand from one chip generation alone. US grid capacity growth is flatlined. Blackwell multiplies the demand again. Tech Buzz China ran the geopolitical companion: "AI's Bottleneck Is Power. The US and China Feel It Differently." Matthew Klein at The Overshoot ran the macro adjacent piece on Europe's physical capacity sitting in the north, productive capacity in the center, demographic capacity in the south, with little overlap.

Week two was the equity-level translation. Stonebridge followed with "Vertiv (VRT): Quantifying the Thermal Moat." With H100s drawing 700W and Blackwell pushing toward 1000W per chip, you cannot cool these densities with standard air conditioning. The Backlog Coverage Ratio was the metric they pushed as the real revenue-visibility tell, against Schneider Electric, Eaton, and Asetek. The trade is not the chip; it is the bottleneck after the chip.

Week three completed the three-week basket with Modine. Stonebridge argued the 100-year-old auto-radiator company is quietly pivoting to hyperscale data-center liquid cooling and is mispriced because screening tools still classify it as Auto Components. By week four, Stonebridge wrapped its "Thermal Moat" basket on Vertiv and Modine as AI-data-center cooling plays. The frame that started as a contrarian napkin in week one had become a three-name thematic basket by month-end. Whether the trades work out is one question; the framing that physical constraint is the 2026 story hiding inside the AI story is the durable contribution.

Arc: Permission You Already Have

Across weeks one and two, three writers who do not share a beat ran the same play on agency and building. The cluster did not extend to the year-end weeks, which makes it the cleanest two-week arc of the month rather than a four-week one. That is its own data point about how the holidays distort attention.

Week one was Furmansky-Kothari-Pupius on the architectural layer. All three argued that the gap between strong and weak operators widened in 2025 rather than narrowed, because AI rewards judgment and punishes the people who use it as a substitute for it. Emily Kramer at MKT1 ran the marketing-side variant on the "Gen Marketer" skillset, generalists fluent in generative AI hired into reshaped org charts. Sahar Mor at AI Tidbits, hosting Jeff Morhous, ran the practitioner companion on Claude Code's subagents, skills, and context files as the operational primitives.

Week two was Kothari, Wee, and Werdelin running the same play from the founder seat. Nikunj Kothari hit it twice. In "Get Your Hands Dirty" he called out the question VCs hear after they ship something with Opus 4.5: "what prompt did you use?" His read: "Most people wait for the job description to change before they change what they do. They're asking themselves for permission they already have." His "Play Rigged Games" borrowed Chris Sacca's frame and applied it to founders pivoting on a market scan instead of lived expertise. Celine Wee ran "Escaping The Swirl," with the diagnosis that the breakdown is not communication, it is that nobody has done the writing. Henrik Werdelin hit the founder version with the A-G-R framing (Agitation, Gravity, Resourcefulness) as the three traits that predict speed better than credentials or ideas. Three writers, three vantage points, one argument: the asset that compounds in 2026 is the willingness to start before the permission lands.

Arc: Year-End Reflection Done Well, and Done Badly

The dominant form of the back half of the month was the year-end retrospective, and the spread between strong versions and boilerplate ones was wider than the genre usually allows.

Week three set the bar high. Nikunj Kothari at Balancing Act wrote the standout with "Plugged into the Matrix," a confessional on what two years of being extremely online has done to his head. He is 35, has two kids, and admits much of his writing comes from anger at people wasting their time. His resolution for 2026: less dopamine, more depth, back to long form. Ben James at Ben by Fax ran the inventory version: a live 3D Tube map, a prototype solar data center, 75 Feathers McGraw replicas around London, a Claude-terminal typewriter, and a fractured jaw from a bike fall. Ben Cmejla at TheGP ran the most operational year-end, a quarter-by-quarter walkthrough of TheGP's bi-weekly AI debates with predictions that look obvious only in hindsight: Figma's David Kossnick called the moment "the Telnet days of AI interfaces" in January, Charlie Labs' Riley Tomasek predicted the death of editor-based coding in February, Alec Flett called that "MCP-first" would soon mean what "API-first" used to.

Week four sharpened the bar further. Nikhil Basu Trivedi at next big thing did the version with teeth, pulling his seven 2025 predictions from a year ago and grading them honestly. AI-Native Applications got an A, with his observation that we have "almost become numb" to twenty-person teams hitting $10M ARR being the line that stuck. Paul Stansik at Hello Operator did the other strong version: pulling his Substack stats and asking which of his own pieces deserved a second read. Most Liked was "What My Dad Taught Me About Work," which he says took longer to write than anything he has published.

The pattern across both weeks was consistent. The retrospectives that worked were the ones where the writer had to confront something they could not gracefully edit around. The ones that curated highlights without grading them faded fast. The other honest version was the writers who admitted out loud that this was a low-output week: Steph Mui saying her newsletter would be shorter for two weeks because she wanted less screen time, Liz Prueitt at Have Your Cake announcing she was taking January off after eighteen months of weekly posting to "reset," The All American running a one-paragraph "we're on break, see you in January" note. The genre rewards specificity and rewards saying nothing when there is nothing to say. The middle ground, the dutiful holiday wrap-up, is where the genre dies.

Arc: The Macro Most People Were Not Pricing

The month's macro thread was small but unusually coherent, and it ran against the rate-cut consensus that had taken over after the December FOMC.

Week two was the Fed read. Matthew Klein at The Overshoot ran "The Fed Submits?", a careful piece on the December SEP showing the median Fed official now expects lower short-term rates through 2027 despite being more optimistic on growth, less worried on unemployment, and just as worried on inflation as in June. His read: the shift cannot be explained by the data, because the government shutdown means most numbers run only through September. The plausible explanation is that a growing cadre inside the Fed is reinforcing administration pressure, while the reserve bank presidents are fighting a rearguard action. The median is misleading; the voting blocs are the story.

Week three was the wage read. Klein returned with the cleanest macro piece of the month: typical American worker wages are still rising at about 4% annualized, versus ~3% in 2018-2019, enough to explain why inflation is running about 1 point hotter than pre-pandemic. The BLS lost six weeks of data collection to the October shutdown, the job market still looks healthy, and the case for rate cuts is weaker than the consensus assumes. Citrini ran "26 Trades for 2026" as the thematic-watchlist companion. Craig Kennedy at Navigating Russia flagged the geopolitical signal: new US, UK, and EU sanctions on Russian oil are the most significant since 2022, with Russia possibly forced to shut in 1.6 to 2.8 million barrels per day, the worst energy-sector crisis there since the 1990s. Tech Buzz China ran Temu Watch 10 and 11 on cross-border ecommerce mechanics and the EU pulling its 150 euro de minimis threshold forward from 2028 to 2026.

The arc here is shorter than the AI ones but the asymmetry is the point. Klein was the only writer all month running a careful argument against the rate-cut consensus. By the time the wage read landed in week three, the bond market was already partially with him. The Fed-submits framing and the sticky-wage framing are the same argument at two altitudes, and both are the right lens for Q1.

The Story of the Month

The story of December is the bottleneck moving from execution to governance, named by Pupius, anchored by Furmansky, validated by Yue Zhao and Mehdi Ouazza from opposite ends of the table, and finally extended into payments infrastructure by Dwayne Gefferie in week four. The whole month is one continuous argument about where the scarce input now sits when the technical work gets cheap.

This was the year of the agent coronation in the press. December was the month operators stopped writing about agent capability and started writing about agent operation, and the literature finally caught up to the tooling. The four-week arc from Furmansky's prompt-band-aid critique to Alephic's context-engineering podcast is the cleanest evolution of operator thinking on AI of any month in 2025. Pair that with Stonebridge's three-week thermal-moat basket and Klein's two-week argument against the rate-cut consensus, and the macro setup heading into 2026 is clearer than the year-end-prediction industrial complex made it look. The model layer is commoditizing while the capital and the power demand are unprecedented. The scarce inputs have moved up the stack to judgment, alignment, and the willingness to surface work in progress. December was the month the operator class agreed on that out loud.

In Retrospect

The Snowflake versus Databricks debate aged poorly inside a single week. SeattleDataGuy's week-two piece "Snowflake vs Databricks Is the Wrong Debate" reframed what most analysts were treating as a vendor question into a workflow consolidation question, with Databricks partnering with Alex the Analyst as the tell. The Pupius and Cmejla pieces in week three then made the broader point that the org chart, not the vendor choice, is where the AI investment thesis actually plays out. Anyone still running a Snowflake-versus-Databricks comparison in December was, in retrospect, looking at the wrong axis.

The Graphite-Warp podcast dropped on the same day Graphite was acquired by Cursor. Taylor Majewski's General Podcast episode with Merrill Lutsky of Graphite and Zach Lloyd of Warp pinned down why code review is becoming the real bottleneck as agents generate more code. The Cursor acquisition the same morning made the episode read as a retrospective rather than a forecast. The framing that code review is the next bottleneck is durable; the assumption that the independent code-review companies would remain independent did not survive the news cycle.

The "consensus rate cut" framing of the December FOMC was the cleanest example of narrative running ahead of data. Klein's "The Fed Submits?" essay was the only major newsletter writing carefully against the grain. By the time his wage read landed the following week, the equity market was still pricing rate cuts and the bond market was starting to price the sticky-inflation case. Most year-end outlooks anchored on the consensus rather than on Klein's vote-bloc framing. The Q1 reading list should weight Klein heavier than the outlook decks.

The 2025-most-clicked round-ups from the marketing newsletters did not age into anything useful. Kyle Poyar polled 130 readers on their top growth experiment in week three; outbound/ABM and partner/ecosystem tied at 20%, well ahead of anything AI-specific. The result is interesting but the format treats marketing as a tactic-of-the-year contest rather than the architectural conversation Furmansky, Kothari, and Pupius were running on the engineering side. The marketing class did not produce its own December equivalent of the Pupius bottleneck-moved essay. That gap is its own data point.

What to Carry Into Next Month

The bottleneck has moved, and the smart move in January is to read every operator essay through that lens. Pupius named it. Yue Zhao named it. Mehdi Ouazza named it. The implication for compensation, for hiring, for org design, and for which 2026 budget lines get cut is direct. Anyone still hiring AI engineers to write better prompts is hiring for the wrong skill. Anyone still buying tools without changing the org chart is, per Kothari, running transformation theater. The right Q1 question is not "which model" or "which vendor"; it is "where in our workflow does the human judgment sit, and how do we instrument for it."

The macro setup is the second frame to carry. Klein on sticky wages plus Stonebridge on the thermal moat is the trade that survives both rate-cut disappointment and AI-capex re-rating. The grid is the constraint nobody on the software side wants to price. The wage data is the constraint nobody on the rate-cut side wants to price. Both will be louder in January. Citrini's "26 Trades for 2026" framing of the outlook format as a watchlist on blind spots rather than a top-call exercise is the right disposition for the next month of reading.

The voice anchor for January reading is the writers who, in week four, admitted what week it was. If you only revisit three pieces from December, Daniel Pupius on going AI native is the cleanest statement of what changes inside a team when code stops being the bottleneck, Matthew Klein's "The Fed Submits?" is the cleanest macro read against the consensus, and Abby Falik's "Why do you walk so fast?" is the post most worth reading slowly when the haze finally lets you. Pupius for the operator frame, Klein for the macro frame, Falik for the pace. Those are the three I am carrying into January.