The Investor

Start with the cash, because that is where the outcome was decided. Equity value near $2.25B against a $1.285B enterprise value implies about $965M of net cash; The Information Dealmaker's version of the same arithmetic uses roughly $900M unspent and ~$2.19B of distributable value. The 7% gap between them changes nothing: roughly 43% of the headline exit is investors receiving back capital they raised and never deployed. Against the ~$1.3B Howie Liu raised over fourteen years, the preference stack got about 1x, and Bain Capital Ventures' Aaref Hilaly puts $850M reaching common.

Then the split. CRV and Caffeinated Capital book sizable returns. XN, Thrive and Greenoaks get near return of capital on five years of duration, in a window where the S&P 500 doubled from September 2020 (the opportunity cost nobody puts in a press release). About 900 employees share $127M, an average near $141K and heavily skewed. Same asset, same buyer, same day; the only variable was entry price.

What the press release left out

Days before signing, per an SEC filing, Airtable carved HyperAgent, its agent build-and-deploy platform launched roughly six months earlier, into a standalone entity, HyperAgent Inc. Existing investors rolled in; Liu runs it full-time post-close and has told them it can raise independently. Neither company commented, and the cost basis on those stakes could not be learned.

Read commercially, a template for stranded 2021-vintage assets: sell the decelerating core to a cash-flow harvester, carve the AI asset into a founder-led vehicle, roll holders in, finance it separately. Also where the objection risk sits. A carveout priced days before signing, with no disclosed independent valuation, is what disgruntled common holders litigate.

A category leader inside 80% of the Fortune 100, still growing over 20%, cleared at a multiple that three years ago described a declining on-prem business.

Where sources agree and diverge

All three accounts agree the buyer matters as much as the price. Bending Spoons is Milan-based, Nasdaq-listed since July 1, 2026, and this is its first post-IPO acquisition; it says it intends to hold long-term and add AI rather than run a sponsor clock. Structurally different from a Vista or Thoma Bravo bid, and for $150M-$500M ARR assets with no realistic listing path, a new liquidity venue.

They diverge on generalizability. The Information Briefing calls Airtable the definitive terminal comp for 2021-vintage horizontal productivity software. Techpresso allows the counter-reading that it is idiosyncratic, a horizontal tool caught between databases and agents, while noting how many 2021 balance sheets share the cash-heavy shape. Tiebreaker: Airtable ranked No. 2 on The Information's enterprise software acquisition-target list, implying a queue of repriced processes anchored here.

What this changes for the book

Two disclosures now carry the weight a revenue multiple used to. Net cash versus operating value, since the treasury clears at par and the business does not. And the full preference waterfall, since a headline mark says nothing about what common receives at a 3x-revenue clearing price. Where the preference stack exceeds a 3x-revenue exit and net cash covers under twelve months, common is wiped out in this scenario, management incentive equity included. Retention breaks before the process starts.

The detail that keeps this from being a short is that Coinbase never stopped buying. Claude Code remains the most-used coding tool among the roughly 2,500 Coinbase engineers who use one, ahead of both Cursor and Coinbase's own internal Forge, and it was chosen by engineers rather than mandated by procurement. Not churn, then. Pricing-power erosion, which shows up in net dollar expansion rather than logo loss, and a portfolio company can bury that in blended metrics for two more quarters.

Five teams built the same product

Coinbase's Forge, Shopify's River (in use by 75% of employees as of a May 2026 presentation), Walmart's Code Puppy, Globant's internal agent and Ramp's own build all do one thing: route requests across Anthropic, OpenAI, Google and open weights behind a single internal interface, distributed through Slack and GitHub rather than the IDE. Five uncoordinated teams shipping the identical missing primitive is category formation, or rather the more useful version of that claim, which is evidence about where the control point moved.

Building in-house relocates the cost rather than removing it. Walmart had to cap internal usage of Code Puppy after demand soared, per Bloomberg's June 2026 reporting, and Coinbase concedes spend ballooned during a stretch with effectively no limits. That is the demand statement for AI FinOps: token attribution, per-team quotas, model-tier policy, cost observability. ROI already benchmarked at roughly half of spend.

The price floor is now set from outside

The supply side points the same way. OpenAI cut GPT-5.6 Luna pricing by 80% to $0.20/$1.20 per million tokens and Terra by 20%, which practitioners read as prior-frontier capability at about 8% of its cost four months ago. DeepSeek V4 Flash ships at $0.14/$0.28 with a 1M-token context, roughly 4.3x cheaper than Luna on output, and Qwen3.8-Max's 2.4T-parameter weights are due imminently. Defensive, not generous. Coinbase experimenting publicly with GLM 5.2 and Kimi K2.7 inside a US-listed, heavily regulated financial institution puts a hard anchor under Western frontier pricing regardless of capability lead.

Token growth stopped being revenue growth, and the buyers built version one of the replacement themselves, badly, at enormous internal cost.

Where the sources disagree

Coverage splits on who keeps the surplus. One read treats the cut as a straightforward COGS gift to the application layer, with analysts expecting savings reinvested into more deployment rather than harvested as CIO budget cuts. The other has anyone billing per token or per message handing the 80% back at renewal, making pricing architecture rather than technology the sorting variable. Both can be true: a company on fixed or outcome-based pricing with 40% inference COGS quietly became a very high gross margin business. The passthrough companies did not.

Three identifiable places this breaks. If one frontier model pulls clearly ahead on agentic coding, routing costs more in rework than it saves and pricing power snaps back. If internal builds stall at 15-20% of workload, the revenue impact on the labs is a rounding error. And any regulatory restriction on Chinese open weights removes the price anchor overnight. Erosion stays the base case.

The financing pattern is the more revealing document here, rather more revealing than the price collapse. Base Power has raised two $1B rounds in roughly ten months against a deployed fleet of just over 500 MWh (about 12,700 homes at 39.2 kWh per unit), with a factory targeting 4 GWh/year and 10+ GWh at the next site. The ~$19/month membership across that fleet comes to something on the order of $3M of recurring revenue. So the subscription is decoration on a merchant energy trading book with a customer acquisition operation and a factory attached. Equity is funding a balance sheet, which is infrastructure capital wearing venture clothing.

The disclosure that actually matters

Utility partnerships went from under 5% to more than half of sales volume in twelve months, with Austin Energy at 40 MW and CoServ, the third-largest US electric co-op, at 100 MW. Read that as the move from volatile merchant arbitrage toward contracted capacity revenue, which is what makes a distributed battery fleet debt-financeable instead of perpetually equity-financeable. It also loosens the geographic ceiling: only about a dozen states plus roughly 80% of Texas have deregulated retail, but co-ops and municipals are a side door into regulated territory.

The edge underneath all of it is regulatory rather than technical, and that is the part worth arguing about. Roughly 2,600 GW of generation and storage sits in US interconnection queues against a grid with 1,279 GW of existing capacity, per Berkeley Lab in June 2026, with connection timelines stretched from under two years in 2008 to nearly five by 2023. There is a reading in which the queue is a temporary bottleneck and the advantage compresses as it clears; nothing in that timeline suggests it is clearing. A homeowner's garage is already interconnected. Base gets the site and the grid connection free while a utility-scale developer waits years for the same megawatts.

Ninety-nine percent of grid storage is still utility-scale, and that is the size of the mispricing the interconnection queue creates.

What has to change in the models

Three inputs are stale rather than conservative. Top-decile hour capture belongs near $0.30/kWh, not $4/kWh, with further compression as fleet capacity grows. The cost stack moved as well: US utilities now invest more in delivery than in generation while the average distribution line carries under half its rated capacity, which is why a behind-the-meter asset competing against retail prices is a different animal from a solar farm competing against wholesale. And 2025 retail prices rose at more than twice inflation, which is the consumer-backlash variable regulators respond to.

The ungated risk sits upstream, and a16z's own framing concedes it: China holds over 80% of battery cell production and over 80% of every stage of solar manufacturing, while Washington's answer is tariffs that raise input costs. Base's Austin plant is pack assembly and integration. Cells are the exposure, and cell origin, tariff pass-through language and a qualified second source are the disclosures worth reading in any hardware energy deal.

This is probably right and possibly overstated: it rests on a single detailed account, and the July 2026 price observation is a brief touch during record load rather than a settled clearing average. The direction is well evidenced. The magnitude deserves independent verification before it enters a model.