Over the next few years, the effective price you pay for high-quality analysis from frontier models will fall faster than the cost of computation ever did under Moore’s Law. This will make entirely new research behaviours economically viable, in the same way cheap storage made digital photos normal, and cheap bandwidth made Netflix even possible. This shift will fundamentally transform AI in investment research, making entirely new workflows economically viable.
How can investors build a research process for that world?
When the unit cost of a technology collapses
History gives us some useful examples.
Storage
In the early 2000’s hard drives cost around US$12 per gigabyte. Today, large drives are down to a few cents per GB, a ~100x+ decline over two decades.
That changed behaviour: We stopped deleting emails. We started versioning everything. Entire business models appeared: cloud backup, photo libraries, “store everything forever” architectures.
If you’d designed your systems in 2000 assuming “storage is expensive”, you’d have made the wrong trade-offs only a few years later.
Bandwidth
In the mid-1990s, the typical home connection was dial-up at ~56 kbps. At those speeds, a single compressed song could take half an hour to download, and the network was effectively text-only.
Today, even mid-tier broadband comfortably streams multiple HD or 4K video feeds.
That bandwidth curve created Netflix, YouTube and Spotify. Zoom and Teams meetings became a no-cost option for everyone. Cloud software that assumes “real-time sync” rather than “download a file, edit, re-upload”.
Netflix: a business model that literally could not exist
In 1995: Consumer internet was mostly 28.8 to 56 kbps dial-up. Home storage was measured in megabytes or a small number of gigabytes. A single 4K movie needs around 50GB of storage. ISPs billed in ways that made constant video streaming uneconomic. A single 4K movie would take at least 3 months (running 24/7) to download.
Netflix was founded in 1997 as a DVD-by-mail company. It didn’t launch streaming until 2007 but even in that year, streaming was a very small part of the business.
Today, Netflix has around 400 million subscribers, who stream on average 90GB of data per month. A global, on-demand video library was physically and economically impossible. It only became possible once bandwidth and storage crossed certain price and performance thresholds.
When a capability gets 10 to 1,000x cheaper, you get entirely new categories of behaviour.
AI is currently on an even steeper curve
Sam Altman recently described what’s happening with AI like this: the cost to use a given level of AI falls about 10x every 12 months, and between early 2023 (GPT-4) and mid-2024 (GPT-4o), the price per token fell by about 150x at similar quality.
You can see this in public pricing. GPT-5 now costs only a few dollars per million tokens, roughly an order of magnitude cheaper than early GPT-4, while also being faster and more capable. Other frontier models (Anthropic, Gemini) have followed suit.
Compared to Moore’s Law, a 2x improvement roughly every 18–24 months, the machine intelligence curve is currently far more aggressive.
For investment teams, a simple working assumption should be: The effective cost per unit of “frontier-level analysis” is on track to fall by an order of magnitude each year, while quality continues to improve.
The ARC-AGI-1 Leaderboard shows this current trend. As the dots on this chart move up and left, users are getting stronger intelligence (a higher % score) for lower cost (US$, logarithmic scale).
Designing a research process for that world means assuming abundance, not scarcity, of machine intelligence.
What becomes possible for AI in investment research??
Let’s consider three time horizons.
1. What’s possible today
With current prices and capabilities, a frontier-model-enabled platform can already support:
Document-centric intelligence: Automatic recording, transcription and summarisation of every company call, roadshow and conference. Structured extraction of KPIs, guidance changes, and risk flags directly into your research system. Review and summarisation of company filings, annual reports and sustainability reports.
Document-rich “Intelligent Notes”: A single research note that can “see” company filings, broker notes, internal models, ESG reports and engagement history, and answer questions like: “How has management changed guidance language on margins over the last four calls?” “Where does this ISS recommendation conflict with our own voting policy?”
Contextual copilots: Per-company or per-portfolio assistants that can draft pre-meeting briefs, first-cut theses, or valuation sanity checks, using both public data and your proprietary research.
At today’s prices, you still have to be somewhat deliberate. You won’t run a full GPT-5 class deep dive on every stock in the global universe every night. But you can comfortably treat the model as a junior associate you can tap many times per name, per week.
2. What’s likely in the near term (12 to 24 months)
If the “10x per year” cost trend holds even approximately, the economics change significantly. You can start to assume:
Always-on surveillance of the investable universe: Continuously scanning filings, news, transcripts and alternative data for your coverage list. Daily “what changed that matters to your thesis?” AI assessment per name, driven off your investment thesis drivers and internal research.
“Shadow coverage” at scale: AI research coverage for thousands of companies, including the entire investible universe outside the primary shortlist, used as a rough prior. Human analysts then go deep where the AI flags mis-alignment, risk or opportunity.
Scenario canvassing instead of one-off memos: For each major position, spinning up rich, branching scenario documents: bull, bear, variant perception, and risk trees, and refreshing them whenever new data arrives.
At this point, we move beyond “what analysis can we afford to run?” and our machine intelligence has progressed to a senior associate rather than a junior.
3. The next few years (2 to 3 year view)
Pushing this out a little further investors should plan for:
Continuous re-underwriting by default: Every holding, every day, is re-evaluated against your investment thesis: quality, valuation, ESG, stewardship, portfolio position sizing. The system flags where the current position and original investment thesis is now inconsistent with the updated evidence.
Agentic research workflows: Autonomous agents that don’t just summarise but act: Pull new data, Update internal scorecards, Draft research notes, Nudge owners of names when something crosses a materiality threshold.
Highly personalised decision support: Portfolio managers receive views and digests tuned to their style: factor biases, risk appetite, preferred evidence. Information is delivered in a format they desire: email, text message, visual dashboard, phone call. The same underlying information is re-shaped differently for a growth PM vs a value PM vs an ESG lead and tailored to their role.
Just as Netflix in 2007 assumed cheaper bandwidth and storage over time, your future research stack should assume cheaper and higher quality analysis is likely.
Designing your process for collapsing cost of intelligence
So how do you take advantage of this as an investment team?
A few design principles we see emerging in platforms like CalibreRMS:
Make the investment thesis the core: Every research artifact – meeting notes, models, result analysis, ESG scorecard – should be used to evaluate the investment thesis it supports. That makes it easy for AI to continuously re-evaluate: “Does our original investment thesis still hold and would we still make this decision today, given all subsequent information and analysis?”
Separate your investment process from model providers: Treat OpenAI, Anthropic, Gemini, Azure OpenAI, Perplexity etc. as interchangeable intelligence engines within your workflow so you can stay at the cutting edge of capability and cost. Your investment process, thesis structures, scorecards and portfolio positions live in your research platform; models are “plug-ins” that can be swapped as prices and capabilities move.
Assume you’ll be doing “too much” analysis and refine the AI to human attention layer: When intelligence is cheap, the bottleneck becomes attention, not computation. Build filters, alerts and dashboards that surface the few situations where the AI thinks your current thesis and the evidence are diverging most.
Exploit the edge of proprietary data: Everyone will have access to the same frontier models running on the same public data. There will be no durable investment edge here. The edge comes from how deeply those models are integrated with your internal process. Proprietary research notes, industry 1/1 meetings, company models, engagement history, voting decisions – and how that data is organised, structured and integrated.
Plan for capability upgrades as a normal part of operations: Just as you upgrade index data or pricing feeds, plan for regular upgrades of your AI: new models, bigger context windows, better tool integration. Your research management system should be able to adopt these without breaking workflows.
CalibreRMS is your intelligent research platform
For a platform like CalibreRMS, whose job is to structure the unstructured and embed research into a single system of record, collapsing model costs are a powerful accelerant for AI in investment research..
Today, all your unique research and company access lives in CalibreRMS and CalibreRMS Intelligence can already summarise, compare, extract and structure across filings, notes, models and meetings.
In the near term, CalibreRMS Intelligence will ingest vastly more sources of information, autonomously evaluate new information, process through intelligent templates, and evaluate the validity of the current investment thesis.
Over the next few years, as the cost of frontier-level intelligence keeps dropping, it becomes realistic to give every analyst, PM and steward their own “always-on” research agent, fully wired into the team’s proprietary corpus and working 24/7.
The CalibreRMS roadmap is designed for a world in which intelligence is abundant and cheap, but attention, insight and judgement remain scarce. The platforms and investment teams that win will be the ones that treat AI not as a gadget, but as a way to help humans make better investment decisions.
——————————
Click Here for Interactive Demos:
Share on Facebook
Share on X
Share on LinkedIn
