A Google engineer used a developer-focused keynote on May 19, 2026 to argue that AI is about to multiply software output by an order of magnitude or more, and that the systems built to handle today's volumes will not hold. The warning has direct consequences for the ad tech and marketing platforms that practitioners depend on every day.

Most coverage of Google I/O 2026 has tracked the consumer-facing announcements: a new model inside AI Mode, persistent background agents, a rebuilt search box, and the Universal Cart shopping hub that spans Search, YouTube, Gmail, and the Gemini app. One session aimed squarely at software engineers received far less attention, yet its argument reaches further into the marketing technology stack than any product reveal.

The talk, titled "Software engineering at the tipping point," was delivered by Adam Bender as part of the Professional Development track at Google I/O 2026. It was livestreamed on May 19, 2026 by the Google for Developers channel and recorded just over 7,600 views in its first day online. Bender, who works on Google's internal developer infrastructure, framed the session around a single question that he said every engineering organisation now has to answer quickly.

A 10x problem stated plainly

Bender's central claim is blunt. "I would bet very good money what we're doing today doesn't work at 10x," he told the audience, adding that the way software is currently built and shipped "doesn't work at 10X velocity."

The figure is not casual. Bender returned repeatedly to multipliers of 10x, 100x, and beyond, describing the coming change as something measured "in orders of magnitude." He framed the timeline tightly. "Suddenly the question of 10x growth is not a thought exercise but a code red moment for you and your company," he said, predicting that organisations "will have to figure this out, if not today but certainly in the next 12 months."

The reason this matters to marketers is structural rather than abstract. Every demand-side platform, every measurement tool, every campaign management interface, and every analytics product a marketing team touches is itself software, produced by an engineering organisation. If Bender is correct that the systems producing software cannot absorb a tenfold increase in output without breaking somewhere, then the tools that marketing departments rely on are exposed to the same pressure. The vendor roadmap, the release cadence, and the reliability of those products all sit downstream of the problem he described.

Why software output is rising, not the value

Bender drew a distinction that complicates the optimistic reading of AI coding tools. "There's a big difference in generating code 10 times faster and generating engineering 10 times faster," he said. The machine that writes code can be made to run fast. Turning that raw output into working systems that deliver results is a separate problem, and one he said the industry has not solved.

He invoked an older idea to make the point. Citing a line attributed to programmer Jeff Atwood, Bender reminded the audience that "software is a liability." More code is not automatically more value. "Right off the bat we may be creating 10 times greater liability if we aren't careful," he said.

That framing has an analogue in the marketing world. Generative tools can already produce campaign copy, creative variants, audience segments, and reports at a pace no human team could match. The volume is real. Whether that volume converts into better outcomes, or simply into more material that someone has to review, validate, and maintain, is the open question. Bender's argument is that volume without judgment accumulates as risk.

The systems that buckle first

The session walked through a simplified model of a software development pipeline and examined what happens to each component when output rises tenfold. The list of failure points was long.

Build systems slow down, because more code means longer compile times and more frequent compiles when agents drive the work. Testing infrastructure strains under load, with Bender noting that a codebase does not grow linearly, so testing everything to confirm nothing breaks could mean running 100 or even 1,000 times as many tests, eventually appearing as a budget line item. Version control systems, he pointed out, are typically optimised for consistency and ordering rather than raw speed, and were never designed for 10x velocity. Code review becomes a human bottleneck, because reviewers cannot sustain the velocity needed to clear the output of even a handful of accelerated developers, and Bender warned that people who do not want to be blockers will start cutting corners in the review process.

He flagged a quieter problem too. If team members are no longer writing code themselves, the review stage becomes their only contact with it, and attention there is thin. "Who is paying attention to the code as it evolves? No one," Bender said. "Pretty soon it will be a mess so no one can understand."

For marketers, the relevant translation is that the platforms they buy could see slower release cycles, more frequent reliability issues, or larger and riskier feature drops, depending on which part of the vendor's pipeline gives way first. Bender's phrasing was that "no one gets out of this unscathed. Scale has effects everywhere."

Tokens, cost, and the economics of invisible work

A recurring theme was the cost of the AI itself. Bender described tokens, the unit of computation consumed by language models, as a real and growing expense. He asked the audience to consider what happens if everyone in a company uses ten times or one hundred times more tokens every day, and posed a follow-up question: "do you know where the tokens are going right now?"

He connected this to Jevons paradox, the nineteenth-century observation that greater efficiency in using a resource tends to increase total consumption of it rather than reduce it. Applied to AI, Bender's expectation is that "tokens will end up everywhere in your workflow." He described the dynamic as putting "a cost on previously-hidden work" that was invisible before, and admitted he did not yet know how that would change behaviour.

The same economic idea has already surfaced in marketing analysis. PPC Land has reported on why AI might create more marketing jobs, not fewer, an argument that rests directly on Jevons paradox: if AI makes campaign creation faster and cheaper, total demand for marketing work could rise rather than fall, because lower costs draw in advertisers who could not previously afford the work. Bender's version of the paradox is more cautionary. He pointed out that load-bearing token engines placed at the centre of a workflow create a new failure mode. "What happens if you don't have enough capacity?" he asked. "If someone ran an agent down and you can't roll back, that's a bad thing."

The agentic web reaches the API layer

One section of the talk carried specific weight for anyone building or integrating advertising technology. Bender warned that the arrival of autonomous agents changes how internal systems must be secured.

"All of your APIs suddenly became public," he said, describing the advice he has been giving colleagues. The reasoning is that agents do not respect informal boundaries. "Agents aren't going to negotiate with you," Bender said. "If they can get access to your data, they will do it, I guarantee you." His recommendation was that internal interfaces need the same hardening normally reserved for systems exposed to the open internet, because "agents will find things you probably didn't want them to."

This connects to infrastructure Google itself unveiled at the same event. During the I/O 2026 Developer keynote, Google announced a WebMCP origin trial in Chrome 149, a mechanism that lets websites expose structured tools for AI agents to call directly rather than forcing agents to parse pixels and page layouts. The Model Context Protocol that underpins it is the standard now being adopted across the industry to govern how agents request data and invoke tools. Bender's warning is the security counterpart to that adoption: as the agentic web moves from prototype to production traffic, the surface area that needs protection expands, and ad tech systems handling campaign data, audience information, and budgets sit within that surface.

SEO consultant Marie Haynes made a related observation in PPC Land's coverage of the search changes SEOs did not expect, noting that Google's recent creation of markdown versions of its developer documentation "was not for search reasons" but "primarily to use it in MCPs." The pattern is consistent: the web is being restructured so that agents, not human readers, are the primary consumers.

Validation becomes the bottleneck

Bender spent time on a problem he called the conjunction of booleans. When very large changes run in parallel and every test turns green, a question remains about the actual reliability of the test infrastructure underneath. "What happens when you have a million tests and the actual reliability of the underlying test infrastructure to run a million tests is in question?" he asked. His conclusion was that it "might not be possible to ship software where every Boolean has to be true."

He also raised the limits of integration testing. Asked whether engineers were happy with their integration testing tools, Bender reported that no hands went up on the livestream, and said plainly: "I am not happy and I don't have the tools to do it the way I want to do it now."

For marketing teams, validation is not a remote engineering concern. It is the difference between a measurement platform that reports accurate numbers and one that does not. PPC Land has documented the accuracy gap repeatedly. A WordStream study found that one in five AI responses for PPC strategy contained inaccuracies, and separate research showed that nearly half of marketers encounter AI errors weekly. Bender's point sharpens the stakes: if the engineering discipline of validation cannot keep pace with AI-generated output, the reliability problems marketers already report could become more frequent rather than less.

Rollbacks, releases, and the speed trap

Bender identified a subtle danger in faster release cycles. Rollbacks work today, he explained, "because you release software slightly slower than you can detect a problem in production." If software can be released faster than problems can be detected, that safety mechanism stops functioning as expected. Every rollback then has other changes stacked on top of it, complicating the recovery.

He also addressed release frequency directly. Teams that already release daily are in reasonable shape, he said, but teams that do not will face larger and therefore scarier changes as AI output accumulates. "The code has to go somewhere to get deployed to be valuable," Bender said. The balance point sits "somewhere between releasing every second and not quite a day."

The marketing relevance is that ad platforms ship changes constantly, and those changes alter how campaigns behave. PPC Land has tracked the consequences of rapid platform change in detail, from Google's algorithm updates landing within 72 hours of each other to the difficulty of isolating the cause of any ranking shift during overlapping rollouts. A faster, AI-accelerated release cadence at the platform level could make that attribution problem harder, not easier.

Democratised building and the loss of intuition

Bender treated the idea that "everyone's a builder" with caution. Democratising engineering is appealing, he said, until an organisation realises it has democratised engineering, and that everyone may now be using different tools with no shared data contract.

He was equally direct about expertise. The reason it takes years to develop a senior engineer, he argued, is that judgment and intuition cannot be installed quickly. "When someone new steps into an environment where they have 50 agents at their disposal and no intuition and judgment, what will go wrong?" he asked. "How do I teach ten years of experience in six months? I don't know yet."

This maps onto a tension marketing teams already face. Generative tools lower the barrier to producing campaigns, creative, and analysis, which widens access. But the Mediaocean survey reported by PPC Land found that marketers are betting heavily on AI media while still struggling to implement generative AI inside their own organisations, with lack of internal expertise cited by 31% of respondents. Access without expertise is a recognised gap in marketing, and Bender described the same gap in engineering.

AI as an amplifier, not a fix

The most portable idea in the session was Bender's framing of AI as an amplifier. He attributed the concept to DORA, the research programme studying software delivery performance. Teams with strong fundamentals can point AI's amplification in productive directions, he said. Teams without them get amplified confusion.

"AI doesn't care where it goes, it's going to give you more of it," Bender said. "AI amplification is not directional, it is just magnitude." More code, more tests, more documentation, and also more confusion, all scaled by the same multiplier.

For marketing leaders, the implication is that AI tools will not repair a weak operation. They will enlarge whatever already exists. A team with disciplined measurement, clear governance, and good data will get more of that. A team without those foundations will get more noise. PPC Land's analysis of ten hard truths separating AI advertising hype from working systems reached a compatible conclusion, documenting that the promise of complete automation remains largely theoretical and that platform opacity, attribution challenges, and infrastructure costs persist regardless of how capable the AI layer becomes.

The forecast Bender offered

Bender closed with a prediction and a timeframe. "In 2030 our developer ecosystems today will feel like 2001 does to us now," he said, noting that in 2001 software still shipped on CD-ROMs. He did not claim certainty about the destination. "We don't know what the future will be but we're figuring it out," he said.

He also named the limit that concerns him most. The problem keeping him up at night, he said, is "how do we maintain intellectual control over this as we grow," which he defined as whether humans can still reason about the systems in front of them. "We have been losing this war for the last 15 years," Bender said, suggesting that AI itself might eventually provide tools to understand large systems as whole systems rather than as collections of individual parts.

The session was a developer talk, but its argument is an industry one. The software pipelines Bender described are the same pipelines that produce the advertising and marketing technology the profession runs on. If those pipelines are entering a period of strain, the products built on top of them inherit that strain. Bender's framing was that everything in a system is connected. For marketers, the connection runs from a Google engineering stage in Mountain View directly to the dashboards open on their screens.

Timeline

Summary

Who: Adam Bender, an engineer working on Google's internal developer infrastructure, speaking at Google I/O 2026. The session addresses software engineers, but its argument reaches marketing and ad tech professionals who depend on software products built by engineering organisations.

What: A keynote talk, "Software engineering at the tipping point," arguing that AI is about to increase software output by 10x, 100x, or more, and that current development systems, including build pipelines, testing infrastructure, version control, code review, and security models, will not absorb that increase without failing somewhere. Bender framed AI as an amplifier of magnitude rather than a fix for weak fundamentals, and warned that internal APIs must be hardened as though they were public because autonomous agents will access whatever data they can reach.

When: The session was delivered and livestreamed on May 19, 2026, during Google I/O 2026, and had recorded just over 7,600 views on the Google for Developers channel in its first day online.

Where: Google I/O 2026, Google's annual developer conference, with the talk distributed through the Google for Developers YouTube channel as part of the Professional Development track.

Why: The argument matters to the marketing community because every advertising platform, measurement tool, and campaign management interface is software produced by an engineering pipeline. If those pipelines are entering a period of strain, the reliability, release cadence, security, and validation of marketing technology products are exposed to the same pressure. The talk also reframes the Jevons paradox economics already discussed in marketing analysis, warns that AI-generated volume accumulates as risk rather than value when judgment is absent, and connects directly to the agentic web infrastructure, including WebMCP and the Model Context Protocol, that Google itself unveiled at the same event.

Share this article
The link has been copied!