Presentation: A Few Predicted Talks From QConAI 2030

Meryem Arik · rss · 2026-09-14 source ↗ Reader 🖥️

AI Labor Disruption AI Regulation & Governance (External)AI Unit Economics Citizen Developer Human as Orchestrator of Agents Jevons Paradox Multi-Agent Orchestration Product Engineer (PM/Engineering Role Merge)Token Throughput

As falling per-unit token costs and near-zero-cost software collide with rising use-case demand, non-technical builders, and parallel agent architectures, the decisive challenge for software organizations through 2030 shifts from producing more code to governing who builds it, what it costs, and how humans supervise it -- a shift that will also redefine what makes a good software engineer and, as public backlash grows, force new AI regulation.

practical · mixed — Genre: a speculative predictions/futurist conference talk, self-referentially framed as a set of fictional talks given at 'QCon AI 2030' (the future conference the audience is imagining). It is practical in ultimate aim -- it tells enterprise early-adopter/early-majority practitioners what to prepare for -- but each prediction is built as a short argument from present-day evidence (expository/argumentative), illustrated with personal anecdotes and analogies (narrative). Treat every forward-looking claim as a forecast by a credible practitioner (Doubleword cofounder/CEO, self-hosted AI inference), not as established fact -- the speaker explicitly flags her own record as one correct prediction (2022/2023 GPU-shortage TEDx talk) plus acknowledged, named biases in her error direction (overestimates enterprise speed, underestimates frontier speed).

Structure

Part	Role
Framing and credibility setup: the speaker states her worldview, methodology, and self-assessed prediction biases, and cites one prior correct prediction (2022/2023 TEDx talk on the coming GPU/inference shortage).	context / ethos
Prediction 1 -- Money: token spend keeps rising even as per-token prices fall (Jevons paradox); token spend management/attribution becomes a major 2030 concern.	evidence + prediction
Prediction 2 -- Code & builders: cost of software trends to zero; non-technical employees become builders; IT's role shifts from coding to running a governed self-service platform.	evidence + prediction
Prediction 3 -- Infrastructure: agents, not humans, are increasingly the customer choosing dev-tool vendors, creating implicit buying decisions and vendor lock-in.	evidence + prediction
Prediction 4 -- Application patterns: parallel ('wide') multi-agent architectures will beat linear ('deep') agents on latency, cost and accuracy, and will dominate enterprise agent design by 2030.	argument + prediction
Prediction 5 -- Regulation: currently weak/toothless AI regulation gives way to real regulation after a 2028 US election dominated by anti-AI, anti-billionaire sentiment.	evidence + prediction
Prediction 6 -- Careers: software-engineer career-leveling frameworks shift away from deep tech-stack knowledge toward product thinking and multi-agent orchestration ('conductor') skill; engineers measured by output-to-cost ratio, not raw token use.	argument + prediction
Closing meta-prediction: by 2030 there will be no separate 'QCon AI' conference -- AI will simply be inseparable from mainstream software engineering.	conclusion

Key terms

token spend management
A not-yet-standard enterprise discipline -- analogous to cloud/FinOps cost management -- for tracking, capping and attributing AI inference cost across an organization, which she predicts becomes a major concern and conference-talk topic by 2030.
fixed intelligence point / fixed IQ point
A device for normalizing token-price trends: holding a constant level of model capability fixed and tracking how much cheaper it gets to buy that capability over time (she claims roughly 10x/year), to argue prices are falling even though total spend rises.
deep agents vs. wide agents
Two ways to apply test-time (inference-time) scaling: a 'deep' agent is a single, largely linear agent chaining sequential tool calls; a 'wide' agent fans a request out to many parallel/specialized sub-agents run concurrently. She predicts wide/parallel architectures win on latency, cost and accuracy.
Cambrian explosion (of bespoke applications)
Metaphor for the coming proliferation of hundreds/thousands of narrow, custom-built internal apps once the cost and skill barrier to building software collapses -- replacing today's model of a few shared apps serving many divergent use cases.
conductor
Her term (arrived at independently of, and echoing, similar framings elsewhere) for the emerging senior-engineer role: someone who directs and evaluates several concurrent AI agents (she cites engineers managing 5-10 at once) rather than personally writing code, valued for holding the whole system in view rather than for deep tool-level skill.
tokenmaxxing (reframed) Token Throughput
A metric she borrows from another QCon talk and reframes: rather than maximizing raw token consumption, the thing to maximize is the ratio of delivered output to total cost (salary plus token spend) -- someone spending $20,000/month on tokens can be exactly as good as someone spending $500/month if their output justifies it.

Propositions

explicitdevelopedAI labs are not subsidizing token prices in order to raise them later; their inference businesses are already profitable, and their overall losses come from infrastructure and training spend, not underpriced tokens.conf: high
extends ←Generative AI Is Not Going To Build Your Engineering Team For You · p6 — Where that talk describes the senior engineer's role shifting toward orchestrating/conducting multiple agents, Majors' account of why single-level teams are dysfunctional implies the same logic should apply to agent orchestration: a team (human or human+agent) that is uniformly 'senior-conductor' work, with no range of levels, would plausibly show the same contention and under-documentation pathologies she describes for all-staff human teams.
explicitThe cost of a fixed level of model intelligence ('a fixed IQ point') has fallen roughly 10x year-over-year for the past three years, driven largely by open-source models distilling frontier capability into more efficient architectures.conf: high
explicitDespite falling per-token prices, total token spend keeps rising -- a Jevons paradox -- because cheaper tokens unlock more use cases and each successive AI-application generation (chatbots, RAG, reasoning models, agents, complex agents) is more token-hungry than the last.conf: high
supportsBaseten CEO Tuhin Srivastava on the AI Inference Crunch, Custom Models · p9 — Independent convergence: both speakers (competing inference-platform founders) separately name Jevons paradox as the reason cheaper tokens produce more, not less, total spend.
explicitBy 2030, token spend management and attribution will be a major enterprise concern -- the percentage of headcount budget spent on tokens will rise significantly and become a real pain point.conf: high
supportsHow to Respond to the Coming AI Cost Shock · p9 — Both independently argue AI token spend should be planned and governed at headcount-budget scale rather than treated as an incidental IT line item.
explicitThe cost of producing software is trending toward zero, illustrated by Doubleword rewriting its entire product frontend for about $20 instead of an estimated two weeks of engineering time.conf: high
explicitAs code becomes effectively free and the skill barrier to building drops, any employee -- not just engineers -- will be able to build software, producing a 'Cambrian explosion' of hundreds of bespoke internal applications per organization.conf: high
explicitWidespread non-technical app-building raises unresolved organizational questions -- who is accountable when a tool breaks, how data standards stay consistent, what happens when the person who built a tool leaves -- that the speaker names but does not resolve.conf: high
explicitBy 2030, IT's role shifts from writing an organization's code to running a governed, self-service enablement platform, with IT teams managing thousands of internal apps built by non-engineers.conf: high
extendsLooking back at the Future of Software Engineering unconference · p9 — Arik predicts IT becomes a platform team managing thousands of citizen-developer-built apps; the other piece already works out a more concrete architectural mechanism (self-healing platform foundation with tight, low-friction constraints) for exactly this scenario.
explicitAI coding agents display strong, consistent infrastructure preferences (Supabase, Vercel, Stripe, Resend), and vendor growth is already substantially driven by agent adoption rather than human adoption -- illustrated by Supabase's user count, most of which arrived in the last year.conf: high
explicitAgents are increasingly making organizations' infrastructure and vendor decisions implicitly, without a human evaluating alternatives, creating vendor lock-in and migration problems the business does not consciously control.conf: high
explicitParallel ('wide') multi-agent architectures -- where a request fans out to many concurrent sub-agents -- will become the dominant enterprise agent-design paradigm by 2030, over today's linear ('deep') single-agent designs, on grounds of lower latency, lower cost via specialization, and higher accuracy via independent trial-and-error.conf: high
explicitMost agents currently built in the enterprise are linear ('deep'), not parallel; parallel/'wide' agent architecture has not yet reached enterprise mainstream.conf: high
explicitParallel agent architectures will generate significant new infrastructure problems -- result aggregation, bursty capacity planning, memory bottlenecks, workload scheduling, observability -- that will dominate future QCon-style engineering talks.conf: high
explicitPresent-day AI regulation is weak: the EU AI Act was drafted mainly before generative AI and is barely enforced, and the US rescinded earlier generative-AI guidance under the Trump administration.conf: high
explicitPublic opinion on AI is already net-negative and cross-partisan: per a cited Pew Research study, 50% of Americans are more worried than excited about AI versus only 10% more excited than worried, with opposition voiced by figures spanning Elizabeth Warren and Bernie Sanders to Republicans like Ron DeSantis.conf: high
explicitThe 2028 US presidential election will be substantially shaped by anti-AI and anti-billionaire sentiment, an anti-AI-leaning candidate will win, new AI regulation follows in 2029, and by 2030 organizations will be implementing meaningfully 'toothy' regulation.conf: medium
supportsCould generative AI turn out to be the tech industry's Vietnam? And could public backlash lead AI to a better path? · p1 — Both independently document fast-growing, visible public backlash against AI as the leading indicator behind predictions of political/regulatory consequences.
explicitBy 2030, deep technology/tech-stack expertise will matter less for software-engineer career leveling, while product thinking, requirements-gathering, and the ability to orchestrate multiple simultaneous AI agents ('conductor' skill) will matter more.conf: high
explicitWithin the author's own organization, some engineers already simultaneously manage between 5 and 10 separate AI agents, functioning as team leads and conductors rather than individual coders.conf: high
supportsThe Conductor Developer · p5 — Independent corroboration of both the 'conductor' framing and a similar numeric ceiling: Arik cites engineers managing 5-10 agents; the other piece cites a felt ceiling of 8-10-12 -- close, independently observed ranges.
explicitEngineer performance will increasingly be measured by a ratio of delivered output to total cost (salary plus token spend) rather than by raw token usage -- someone spending $20,000/month on tokens can be exactly as good as someone spending $500/month if their output justifies it.conf: high
explicitBy 2030 there will no longer be a distinct 'QCon AI' conference -- AI practice will simply be inseparable from mainstream software engineering, folded back into general QCon.conf: medium
implicitThe talk implicitly treats the speaker's single prior correct prediction (the 2022/2023 TEDx talk on the coming GPU/inference shortage) as a meaningful signal of the reliability of her present, much broader set of six predictions -- despite the speaker's own caveat ('I was right once, I might not be right again').conf: medium
implicitThe talk implicitly assumes that 'early adopter and early majority' enterprises are the correct reference population for what 2030 will look like, treating adoption-diffusion lag among laggard organizations as out of scope rather than as a variable that could push several of these predictions past 2030.conf: medium
implicitThe parallel-agents argument implicitly assumes the infrastructure problems it creates (aggregation, capacity planning, memory bottlenecks, scheduling) are solvable engineering complexity worth the cost, rather than evidence that the wide-agent paradigm is fundamentally less tractable at enterprise scale by 2030.conf: medium

Arguments

Token spend will become a top enterprise pain point by 2030, with the share of headcount budget spent on tokens rising significantly. strength: medium

Explicit premises:

Token prices for a fixed intelligence level are falling roughly 10x per year
Despite falling prices, aggregate spend is rising (Jevons paradox)
Every successive AI use-case generation (chatbots, RAG, reasoning, agents, complex agents) is more token-hungry than the last
Frontier models released by labs tend to be larger/more expensive even as older models get cheaper
A cited Goldman Sachs estimate projects agent token usage multiplying 24x by 2030, which she considers conservative

Implicit premises:

Enterprises will not proactively govern token consumption ahead of the pain becoming visible, i.e. they under-invest in AI-cost governance until forced to
'Percentage of headcount budget spent on tokens' is a metric executives will actually track and act on, not just an analyst framing device
Growth in new use cases will keep outpacing gains in cost-per-use-case efficiency at least through 2030

Inferences:

If token spend reaches headcount-budget scale, procurement/finance functions -- not just engineering -- will need direct authority over AI inference purchasing.
Token cost-attribution tooling becomes a precondition for any credible chargeback or ROI accounting on AI initiatives.

IT's core function will shift from writing code to operating a governed self-service platform for non-technical builders, managing thousands of internal apps by 2030. strength: medium

Explicit premises:

The cost of producing software is trending toward zero
Non-technical employees can already build working custom apps in minutes with current AI tools
Every organization has far more unmet bespoke use cases than the handful of shared apps it currently funds can serve
Unmanaged proliferation of bespoke apps creates cyber-risk, data-silo, and accountability problems

Implicit premises:

Demand for bespoke software is effectively unbounded once cost and skill barriers disappear -- people always have more unmet workflow needs than they currently express
Organizations will choose to govern this proliferation through platform enablement rather than by banning, ignoring, or failing to contain non-technical building
IT organizations currently possess, or will acquire in time, the platform-engineering capability needed to do this at scale, rather than being overwhelmed by it

Inferences:

Hiring for 'IT' roles shifts toward platform-product-management skill and decreasingly toward the ability to personally write features.
Governance failures (an app's non-technical builder leaving, unowned production dependencies) become a routine, designed-for operational risk category rather than an edge case.

Parallel ('wide') multi-agent architectures will become the dominant enterprise agent paradigm by 2030, displacing today's linear ('deep') agents. strength: low

Explicit premises:

Parallel inference lowers end-to-end latency because only throughput, not a sequential call chain, matters
Parallel/specialized sub-agents can be made cheaper through smart context management and use-case specialization
Running multiple agents on the same task and comparing/voting on results improves accuracy
Frontier agent products she names (Claude, Perplexity) already use parallel sub-agent patterns internally

Implicit premises:

What works at the frontier-lab product level (parallel sub-agents inside a single vendor's product) will transfer to how individual enterprises architect their own agent systems, rather than remaining a vendor-side implementation detail invisible to enterprise builders
The added infrastructure complexity of parallel agents (aggregation, capacity planning, memory, scheduling) is an engineering cost worth paying for the latency/cost/accuracy gains, not a reason enterprises will stay with simpler linear agents
Enterprises will develop the tooling and organizational maturity to operate bursty, 'chaotic' parallel-agent workloads within roughly five years

Inferences:

A new specialized infrastructure/tooling category (agent-result aggregation, agent-capacity autoscaling) becomes investable and buildable in its own right.
AI reliability engineering converges toward classic distributed-systems practice -- aggregation, backpressure, observability -- rather than remaining a discipline apart.

The 2028 US election will produce a meaningfully AI-hostile political and regulatory environment reaching 2030. strength: low

Explicit premises:

Current AI regulation (EU AI Act, prior US guidance) is weak or has been actively rescinded
Public opinion is already net-negative on AI (Pew: 50% more worried vs. 10% more excited)
Cross-partisan political figures (Warren, Sanders, DeSantis) already voice anti-AI positions
AI's economic impact has not yet been broadly felt, which she expects will intensify sentiment as it arrives

Implicit premises:

Current elite/political anti-AI rhetoric will convert into an electorally decisive campaign issue by 2028, rather than staying rhetorical without translating into votes
Whichever administration is elected in 2028 will both want to and be able to pass meaningfully 'toothy' regulation within about a year (2029), despite the historically slow pace of US federal tech regulation
Anti-AI and anti-billionaire sentiment will remain politically fused as a single current through 2030 rather than fracturing along other lines (national security vs. labor vs. safety)

Inferences:

Enterprise AI roadmaps for 2028-2030 should include political/regulatory-risk contingency planning at the same seriousness level as technical or cost risk.
Bipartisan anti-AI-billionaire messaging could combine with existing local anti-data-center movements into a broader electoral coalition.

Deep technology/tech-stack expertise will become less important to software-engineer career leveling by 2030, while product thinking and multi-agent orchestration skill become more important. strength: medium

Explicit premises:

Current career-leveling frameworks reward technology depth and system ownership largely because onboarding someone new onto a complex stack is hard
AI tools make onboarding to and understanding a complex stack much easier, reducing the retention value of deep system knowledge
Companies are hiring markedly more 'product engineers' than before
Engineers increasingly work by directing several concurrent agents rather than collaborating within large human teams

Implicit premises:

Career-leveling frameworks exist primarily as retention tools tied to the scarcity of tacit system knowledge, so once AI erodes that scarcity the frameworks' underlying logic -- not just their surface criteria -- necessarily changes too
Product/requirements-gathering skill cannot itself be substantially automated by AI the way technical execution has been, making it a durable differentiator rather than the next skill AI erodes
The retention/leveling dynamics observed at her own (apparently small, high-agency) startup generalize to large enterprise engineering organizations with very different team structures and entrenched leveling systems

Inferences:

If technical depth stops gating career progression, the traditional IC-to-staff-engineer ladder may need a parallel 'product-engineer' ladder with different promotion criteria.
Junior-engineer onboarding and mentorship, historically built around gradually acquiring deep system knowledge, needs redesigning around a different, less well-understood scarce skill (orchestration and product judgment).

Narratives

Doubleword rewrote its entire product frontend for roughly $20 in AI-generated work instead of an estimated two weeks of engineer time; separately, the speaker used AI to build custom wedding-planning software rather than use a spreadsheet.
→ Makes the abstract claim 'the cost of software is trending toward zero' concrete and personally credible, persuading by demonstration rather than by argument or data.
A non-technical business-development-rep (BDR), fresh out of college, built a working lead-tracking app -- connected to a lead-capture device, HubSpot, and email systems -- in about 25 minutes at a conference.
→ Makes the 'anyone can build software' prediction vivid and personal, and implicitly argues that this capability already exists today, not as a 2030 hypothetical -- strengthening the plausibility of the more speculative 'Cambrian explosion of bespoke apps' prediction that follows.
The speaker's prior TEDx talk (given a couple of months after ChatGPT's launch) argued inference demand would outstrip GPU supply; NVIDIA's stock subsequently rose roughly 10x.
→ An argument from personal authority/track record that substitutes for direct evidence of the present talk's specific predictions being accurate -- she explicitly uses it to justify her 'confidence' in making today's forecasts, while also hedging it ('I was right once, I might not be right again').
She recalls working as a quant at an investment bank where every employee kept their own version of a financial model in a separate Excel spreadsheet -- the same fragmentation problem organizations already lived with before software was cheap.
→ Normalizes the coming proliferation of bespoke AI-built apps as a known organizational pattern recurring in a new, code-based form rather than a genuinely novel problem -- softening the 'Cambrian explosion' prediction's apparent radicalism.

Questions addressed

✅ Will AI labs raise token prices once any early subsidization ends?
No -- she argues inference is already profitable for the labs and open-source competitive pressure will keep pushing effective prices down, not up.
✅ Will overall enterprise AI/token spend go up or down as per-token prices keep falling?
Up, via Jevons paradox -- cheaper tokens unlock more, and more token-hungry, use cases faster than prices fall.
⭕ Who is accountable for infrastructure/vendor decisions once AI agents make them implicitly, without a human evaluating alternatives?
She explicitly names this as an open, unresolved question ('I think this is going to lead to really interesting conversations about how do we avoid vendor lock-in').
✅ Which agent architecture -- linear/'deep' or parallel/'wide' -- will dominate enterprise AI applications?
Parallel/'wide' agents, on latency, cost, and accuracy grounds, despite currently being the minority pattern in enterprise deployments.
✅ Will there be meaningful AI regulation by 2030, given how weak it is today?
Yes -- following a 2028 US election dominated by anti-AI/anti-billionaire sentiment and new regulation drafted in 2029.
✅ If deep technical knowledge matters less, what will actually make a good software engineer in 2030?
Product thinking, requirements-gathering, and the ability to orchestrate ('conduct') several concurrent AI agents; performance measured by an output-to-total-cost ratio rather than raw token use.
⭕ Once IT becomes a 'platform team' enabling thousands of non-technical-built apps, does that framing actually answer the specific accountability, ownership, and cyber-risk questions she herself raised -- or does it just relocate them?
Not directly addressed; she poses the accountability questions as a rhetorical list, then moves to the 'IT becomes a platform team' resolution without walking through how platform constraints actually answer each named risk.

Gaps

incompleteThe 2028-election-to-2029-regulation timeline is asserted with no engagement with historical base rates for how long diffuse public tech backlash typically takes to become fast-moving federal legislation (e.g. financial-crisis-era banking anger to Dodd-Frank, or social-media backlash to any comparable US regulation, which remains largely unenacted years later). The one-year sentiment-to-legislation assumption is unstress-tested.
Read next: Read a history of technology-regulation lag after comparable public backlash episodes to test whether a 2028-election-to-2029-regulation timeline is plausible or optimistic.
incompleteThe 'parallel agents will win' prediction cites only vendor-side evidence (that frontier products like Claude and Perplexity use parallel sub-agent patterns internally) and no cost model or enterprise production case study. This sits in tension with a more cautious, incident-grounded account elsewhere in the corpus that ties the practically workable number of parallel agents to human code-review throughput (a low single digit) rather than to an infrastructure-solvable ceiling.
Read next: extract-osmani-orchestration-tax-is-you and extract-fowler-orchestrators-tax, for grounded, incident-derived accounts of what actually breaks when parallel-agent fan-out scales inside a real team.
incompleteThe accountability and governance questions for non-technical-built software (who owns a tool when its builder leaves, how data standards stay consistent, who is liable for a broken agent-connected app) are posed as a rhetorical list but never actually answered -- 'IT becomes a platform team' is asserted as the resolution without showing how platform constraints resolve each named risk.
Read next: extract-looking-back-at-fose-2026, which works through a comparable citizen-developer/platform-governance problem with more concrete mechanism (self-healing foundations, tight low-friction constraints guiding both builder and agent from the first prompt).
incompleteThe talk addresses how an already-senior engineer's role changes (from coder to 'conductor') but not how anyone becomes qualified to do that role in the first place -- if deep tech-stack knowledge stops being the differentiator and the entry ticket, the pipeline by which a junior engineer earns the judgment needed to evaluate multiple agents' output well is left unaddressed.
Read next: Corpus material on the broken junior-engineer on-ramp (ai-labor-disruption cluster) for the entry-level side of this same shift.
closed byWhen AI Writes Almost All Code, What Happens To Software Engineering? — That gap asks how anyone becomes qualified for the senior 'conductor'-style role if deep tech-stack knowledge stops being the entry ticket. This article answers directly: it proposes rigorous CS/software-engineering education as a near-non-negotiable hiring filter (teaching architecture, testing, and tech-debt concepts before day one) and argues new grads may 'mature faster' by skipping the pure-coding apprenticeship stage and jumping straight to engineering judgment — an explicit, if untested, proposal for the pipeline the other document leaves open.
incompleteDoubleword is itself a commercial self-hosted-inference vendor whose business benefits from enterprises worrying about token cost and wanting more control over inference -- the talk does not acknowledge this potential conflict of interest when making 'token spend will keep rising and become a crisis' predictions that are directly aligned with the speaker's commercial interest.
Read next: Read vendor-neutral AI-inference-cost analyses, and triangulate against other vendor-authored accounts already in the corpus (e.g. extract-baseten-ceo-inference-crunch, itself from a competing inference vendor), to check whether the magnitude of the 'token spend crisis' framing holds up outside vendor messaging.
illogicalThe speaker states up front that she personally tends to overestimate how quickly enterprises change, but does not apply that self-flagged bias when she reaches her most enterprise-speed-dependent predictions -- 'IT manages thousands of apps by 2030' and 'no more separate QCon AI by 2030' both assume fast, broad enterprise organizational change, without her flagging the tension with her own stated error direction.
Read next: Re-read the talk's specific 2030 timelines discounted for the overestimation bias she herself names, to see which predictions would more plausibly land closer to 2032-2035.

Presentation: A Few Predicted Talks From QConAI 2030

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents