AI’s Next Winners? Investor Bets on Snowflake, CrowdStrike and Palantir

The Hook

Today's edition: three enterprise stocks named as AI's next plays, a new framework for training smarter agents from their own logs, and a protein-modeling paper with real drug-discovery stakes. Let's get into it.

The Signal

1. The Infrastructure Bet: Snowflake, CrowdStrike, Palantir
An investor note circulating this weekend named these three as the companies best positioned to capture enterprise AI spending — not as model builders, but as the infrastructure layer where AI gets deployed at scale. The logic: enterprises don't run on raw models. They run on data platforms (Snowflake), security tooling (CrowdStrike), and decision-intelligence systems (Palantir). The note argues all three already sit inside enterprise IT stacks and are expanding AI-driven features on top of existing contracts — no cold-start problem, no new procurement conversation. What this means for you: if you're evaluating AI vendors at work, integration friction with your existing data and security infrastructure will dominate your decision far more than any benchmark score.

2. Teaching Agents From Their Own Mistakes
A new paper, Trace2Tower, introduces a framework for training LLM agents — large language model-based autonomous systems — to build multi-level skills from execution traces. A trace is the step-by-step record of what an agent did: state at time T, action taken, new state at T+1. Current approaches mostly treat these logs as flat replay data. Trace2Tower proposes inducing a hierarchy of skills from those logs — lower-level primitives like 'query a database row' and higher-level composites like 'reconcile two conflicting records' — using a technique called EigenTrace Induction. The key insight is that transitions between states carry more signal than the states themselves. Early benchmark results show significant gains on interactive task evaluations. Practical angle: if you're building agents today, instrument for state transitions, not just final outputs.

3. Protein AI Gets a Memory Upgrade
ProtLingo is a new protein language model — a model trained on amino-acid sequences the way GPT is trained on text — that adds two architectural improvements: conditional memory and expert routing. Conditional memory lets the model selectively retain context from earlier in a long protein sequence, solving the problem where standard transformers lose track of dependencies across hundreds of amino acids. Expert routing — part of a mixture-of-experts architecture — lets different sub-networks specialize on different protein families. The result: stronger prediction of the functional impact of single amino-acid mutations at lower computational cost than prior models. Drug discovery teams use exactly this capability to screen candidate compounds. This is frontier AI being quietly useful where the downstream stakes are genuinely high.

4. The Gold Migration Signal
Several European countries have been physically moving gold reserves out of North American vaults and back onto home soil. The story pulled 194 Hacker News upvotes and over 300 comments this weekend, the strongest organic-interest signal in today's entire story pool. The macro read: de-dollarization pressure is real enough that sovereign governments are acting on it physically. For the AI reader, the connection is indirect but load-bearing — the same geopolitical friction shapes semiconductor export controls, cloud infrastructure geography, and where AI compute gets built and regulated. The physical movement of gold is the most visible symptom of a structural shift the tech industry will be navigating for years.

5. Hormuz Chokepoint
Iran's security chief announced this weekend that Tehran will declare a restricted zone outside the Strait of Hormuz, through which roughly 20 percent of global oil supply passes. If enforced, this raises shipping risk, insurance costs, and energy prices across global supply chains. Data centers are not immune to energy cost shocks — GPU compute is energy-intensive, and cost increases propagate through inference pricing. Not an immediate AI story, but worth one eye as it develops.

Quick Hits

• fastcore 2.2.22 dropped on PyPI this weekend — the utility library underlying the fastai ecosystem got a minor update. If you're building Python ML pipelines or agent tooling on fastai foundations, staying current on fastcore avoids quiet compatibility breaks downstream.
• Alcaraz into the US Open quarters — straight sets over Tommy Paul. The only AI-adjacent angle: Palantir's analytics contracts continue to expand into new sectors., and a high-profile Alcaraz run keeps that use-case visible.
• Mexico festival fireworks blast — at least 10 killed, 60 wounded, triggered by a burning bull effigy. No AI angle. A reminder that the most consequential safety failures are often low-tech, and that real-world harm benchmarks matter when AI safety researchers calibrate risk frameworks.

The Cold Open

It is a Sunday-into-Monday kind of morning. Somewhere, an investor is circling three company names on a note — names that millions of people already own — and calling them AI's next infrastructure winners. Somewhere else, a model is reading amino-acid chains like sentences, predicting what breaks when you change one letter in a protein that determines whether a drug candidate works. Two very different expressions of the same underlying shift: AI moving from demonstration into infrastructure, from benchmark into working system. That tension between financial positioning and genuine technical progress is the story of this moment in AI. Today we look at both ends of it.

The Anchor

Why Snowflake, CrowdStrike, and Palantir — and What It Signals About the Enterprise AI Thesis

The investor note naming these three as AI's next winners is worth unpacking carefully, because the logic it uses tells you something about where value actually accrues in an AI adoption cycle — and it might not be where you expect.

None of the three are model builders. They do not compete with Anthropic, OpenAI, or Google DeepMind. Snowflake is a cloud data platform — its core product is letting enterprises store, query, and transform large datasets without managing their own infrastructure. CrowdStrike is an endpoint security company — it watches every process running on every device in an enterprise network and flags anomalies. Palantir builds decision-intelligence software — it turns messy operational data into structured views that analysts and executives can act on.

What the three share: they already have enterprise contracts. And those contracts give them something more valuable than a model — they give them the data relationship. Snowflake knows what queries your analysts run. CrowdStrike knows what your network traffic looks like at baseline. Palantir knows the shape of your decision workflows. The AI thesis is that each company is now positioned to layer models on top of that existing relationship and sell the AI-augmented product as an upgrade, not a new purchase.

The products already exist. Snowflake's Cortex lets customers run LLM queries against their own data warehouse. CrowdStrike's Charlotte AI assistant surfaces threat intelligence inside the security dashboard analysts already work in. Palantir's AIP platform wires generative AI into the decision workflows that defense and commercial customers rely on. None of these require a new procurement conversation. They ride the existing contract.

There is a counter-argument worth naming directly. If data relationships are the moat, Salesforce and SAP have spent decades embedding themselves in enterprise workflows. The specific bet on Snowflake, CrowdStrike, and Palantir likely reflects one of two things: a view that larger incumbents integrate AI more slowly because they have more legacy to protect, or simple valuation math — the upside multiple on a sixty-billion-dollar company is larger than on a two-trillion-dollar one.

For a working engineer or product person, the takeaway does not require taking a position on the stocks. The underlying insight is practical: when you evaluate AI tooling for your team or your company, integration friction with your existing data and security infrastructure will dominate your decision more than model quality benchmarks. The model that wins inside your organization will not be the model that scores highest on MMLU. It will be the model embedded in the system your data already lives in. That is the enterprise AI thesis, and today's investor note is one more public articulation of it.

Deep Dive

Trace2Tower: How to Teach an Agent From What It Did

The paper's full title — 'Trace2Tower: Transition-Aware EigenTrace Induction of Multi-Level Skills for LLM Agents' — is dense. Let's unpack it layer by layer, because the mechanism is genuinely interesting and the engineering implication is immediately applicable.

The problem it is solving. LLM agents — autonomous systems built on large language models that take sequences of actions to complete a task — currently learn from two main signal sources: human-written demonstrations, which are expensive and don't scale, and outcome-level feedback, like 'task succeeded' or 'task failed.' What neither captures well is the intermediate structure of a complex task — the hierarchy of sub-skills that a competent agent strings together to get from start to finish. A capable human agent doing a research task doesn't just know 'search' and 'report' — they know how to recognize when a search result is ambiguous, shift to a verification sub-task, resolve the ambiguity, and then return to the main task thread. Current training approaches largely ignore that hierarchical structure.

What an execution trace is. When an agent runs, it produces a trace: a timestamped sequence of states and actions. State at time T, action taken, resulting state at time T+1, and so on — a complete flight data recorder for the agent's decision process. Current approaches treat these traces as flat training data: replay the (state, action) pairs, fine-tune the model on the sequence, done. The structural information about which actions cluster into coherent sub-tasks is largely discarded.

The EigenTrace insight. The paper's core technique is called EigenTrace Induction, borrowed from linear algebra. The authors compute a transition matrix over agent states — how often does state A lead to state B across a corpus of traces — and extract the dominant transition patterns using eigenvector decomposition. Those dominant patterns correspond to coherent sub-tasks: the natural 'chapters' of agent behavior that recur across different task instances. The paper calls these induced patterns multi-level skills, organized into a tower: low-level primitive actions at the base, mid-level procedural skills in the middle, high-level compositional strategies at the top.

Why transition-aware matters. Most trace-based learning focuses on individual (state, action) pairs. Transition-aware learning focuses on the moments of behavioral shift — when the agent recognizes that one sub-task is complete and the next has begun. The paper's argument is that this transition signal is more generalizable than action-level signal: the specific keystrokes an agent uses to query a database vary across tasks, but the recognition that a data-retrieval phase has concluded and a synthesis phase has begun is structurally stable.

The engineering implication today. If you are building agents using any current framework — LangChain, LlamaIndex, custom function-calling loops — the insight is immediately applicable without waiting for this paper's approach to ship in a library. Log your agent's state transitions explicitly. Tag each step in the trace with a phase label: 'data retrieval,' 'validation,' 'synthesis,' 'error recovery.' Record the timestamp and agent state at each phase boundary. Even if you are not training a model on these logs today, you are building the annotated dataset that the next generation of agent training approaches will require. Transition-annotated logs cost almost nothing to generate and compound in value as your agent accumulates run history.

One Technique

State-Transition Logging for LLM Agents

If you are building or evaluating an LLM agent this week, add one thing to your instrumentation: explicit state-transition logs. Most teams log inputs, outputs, and errors. Few log the moment an agent shifts from one sub-task phase to another — but that transition moment is precisely where the Trace2Tower paper finds the most reusable skill signal.

In practice: tag each step in your agent's trace with a short phase label (e.g., 'retrieval,' 'validation,' 'synthesis,' 'error-recovery'). Log the timestamp and a snapshot of relevant agent state at each phase boundary. Store these as structured JSON — one log file per agent run, with a phase_transitions array alongside the standard action log.

You do not need to be training a model to make this worthwhile. Transition-annotated logs make debugging faster (you can see exactly where in the task hierarchy an agent went off-track), make evaluation cleaner (you can score sub-task phases independently), and give you ready-made training data the moment you want to improve the agent from its own history. Three lines of logging code now, substantial leverage later.

One Prompt

Use this prompt to extract phase-transition structure from an existing agent log or conversation trace. Paste your agent's run log as context, then run:

You are an agent behavior analyst. I will give you an execution trace from an LLM agent: a sequence of steps, states, and actions. Your job:

1. Identify the natural sub-task boundaries in this trace — the moments where the agent's behavior shifted from one phase to another.
2. Label each phase with a short descriptive name (e.g. 'data retrieval', 'validation', 'synthesis', 'error recovery').
3. For each transition boundary, note: what triggered the shift, and what changed in the agent's approach afterward.
4. Output a structured list: Phase name | Start step | End step | One-sentence description | What triggered the transition.

Here is the trace:
[PASTE AGENT LOG HERE]

Works best with tool-calling agent traces (function calls plus results) or multi-step chain-of-thought logs. The output is immediately usable as a manual annotation pass for transition-based training data, or as a diagnostic view when your agent goes off-track.

One Tip

Check your AI vendor's data residency setting before your next demo.

With European gold repatriation in the news and data-sovereignty pressure accelerating, this is a good week to verify one concrete thing: where does the AI tool you're using actually process and store your data? Most enterprise AI vendors have data residency options — EU-only, US-only, private cloud deployment — that are not enabled by default. If you're demoing a tool to a European customer, or working with any data that touches GDPR scope, check the vendor's data processing agreement before you paste anything into a prompt. This takes five minutes and prevents a compliance conversation you do not want to have retroactively.

Tool of the Day

fastcore — version 2.2.22, available at pypi.org/project/fastcore

fastcore is a Python utility library built by the fastai team that adds typed dispatch, productivity patterns, and convenience functions on top of standard Python. It is the foundation that fastai, nbdev, and related tools are built on.

What it is genuinely good for: if you write Python for ML, data pipelines, or agent tooling, fastcore's typed dispatch system gives you clean polymorphic functions without the boilerplate of standard Python singledispatch. Its delegates pattern simplifies wrapping classes that you do not own. The test utilities catch edge cases with minimal syntax overhead. These are not glamorous features — they are the kind of thing that makes a codebase noticeably cleaner after six months of use.

Honest limits: fastcore is built for the fastai style of Python, which assumes comfort with functional patterns and minimal ceremony. If you are coming from a Java or strongly-typed TypeScript background, some patterns will feel loose. Documentation is sparse outside the fastai ecosystem — the best way to learn it is reading fastai source code directly, which is itself clearly written but requires some orientation time.

Signature Bites

• Enterprise AI follows the data contract, not the benchmark. Where your data already lives is where AI gets deployed first — model quality is secondary.
• Agent traces are training data — log transitions, not just outputs. The shift between sub-tasks carries more reusable signal than the action taken inside one.
• ProtLingo's efficiency matters as much as its accuracy. A mutation-prediction model is only useful to drug discovery if it is fast and cheap enough to screen candidates at scale.
• Geopolitical pressure does not stop at physical assets. Gold repatriation and chip export controls are the same underlying structural story at different altitudes.

Joke of the Day

An LLM agent was asked to plan a shipping route through the Strait of Hormuz. It returned 47 tool calls, a comprehensive geopolitical risk assessment, and a strongly worded recommendation to remain in the data center.

Fact of the Day

The Strait of Hormuz narrows to a tight chokepoint at its most constrained stretch. — yet A significant share of global oil and liquefied natural gas trade passes through that gap every day. It is the single most consequential maritime chokepoint on Earth, and Gulf exporters have no realistic alternative route. A restricted zone announcement there moves energy markets globally within hours.

Stat That Matters

The European gold repatriation story drew notable organic interest on Hacker News this weekend. In a feed dominated by technical AI content, a story about sovereign governments physically moving gold is outperforming everything else. What it signals: macro risk and geopolitical uncertainty are now primary context for how the engineering and tech-investor community thinks about AI infrastructure decisions — not background noise, not a separate conversation.

Trends

Agentic AI accounts for the largest share of today's corpus, ahead of funding and frontier research coverage. The volume confirms what Trace2Tower represents: agent capability is the current active frontier of practical AI development, and the research community is converging on it fast. The funding lane tracking closely behind suggests investor attention is following research momentum with roughly a one-cycle lag. Consumer AI continues to hold a steady presence in today's corpus. — iterative-improvement mode rather than breakthrough mode this week. Policy and security lanes are low in AI-specific volume but structurally elevated by the Hormuz and gold stories, which set the macro backdrop for every infrastructure decision in this space.

Bold Prediction

Within 18 months, at least one of the three companies named in today's investor note — Snowflake, CrowdStrike, or Palantir — will be publicly credited with displacing a standalone AI-native vendor from a named Fortune 500 enterprise contract. The displacement mechanism will not be superior model quality. It will be procurement consolidation: an existing customer choosing to expand the AI feature inside a contract they already have rather than maintain a separate AI-native vendor relationship. The prediction is falsifiable: a named Fortune 500 customer publicly confirms switching from a standalone AI tool to an AI feature inside their existing Snowflake, CrowdStrike, or Palantir deployment. Watch for it in earnings call commentary starting Q1 2027.

Paper Watch

ProtLingo: Efficient Protein Language Modeling via Conditional Memory and Expert Routing
arXiv:2609.04793

Proteins are sequences of amino acids — hundreds to thousands of residues long — and small changes in that sequence can dramatically alter what a protein does in the body. Stability, binding affinity, enzyme activity: all of it can hinge on a single substitution. ProtLingo treats protein sequences the way a language model treats text: as a sequence of tokens with long-range dependencies that must be modeled correctly to understand meaning.

The two improvements it introduces are architectural and practical. Conditional memory solves the problem of a standard transformer losing track of residues it saw 400 positions ago in a long sequence — it selectively retains relevant earlier context rather than compressing everything equally. Expert routing assigns different sub-networks to handle different protein families — the way a specialist outperforms a generalist on their specific domain. The result is improved prediction of single-mutation functional effects. For drug discovery pipelines, this translates directly: more candidate compounds can be screened per dollar of compute, which means more shots on goal in the search for viable therapeutics.

Founder Spotlight

Alex Karp, Palantir Technologies

Palantir's CEO has spent a decade making a bet that looks less contrarian every quarter: that enterprises and governments would pay for AI-augmented decision workflows before they would pay for raw model access. The investor note naming Palantir alongside Snowflake and CrowdStrike is a public validation of that thesis reaching mainstream investor consciousness.

The strategic move worth watching is how Karp positioned AIP — the Palantir AI Platform — not as a model or a chatbot but as a workflow layer that sits between an organization's data and its human decision-makers. That framing is now the standard enterprise AI pitch across the industry. Palantir arrived at it early, when the consensus still assumed the value would accrue to model builders.

The open question going into 2027: does being early to a positioning also mean being sticky once the large platform vendors replicate the workflow-layer concept? Microsoft Copilot, Salesforce Einstein, and SAP's AI offerings are all converging on the same frame. Palantir's defensibility rests on the depth of its operational integration — the degree to which customers have built actual decision processes around its specific tooling. Shallow integration commoditizes; deep integration compounds. That distinction will determine whether today's investor thesis ages well.

Quote

'Enterprises don't run on raw models; they run on data platforms, security tooling, and decision-intelligence systems.'

— Paraphrased from the investor note on Snowflake, CrowdStrike, and Palantir, September 2026

Learner's Edge

What Is a Mixture of Experts (MoE)?

A mixture-of-experts model — MoE — is a neural network architecture where, instead of every part of the network processing every input, different sub-networks called 'experts' specialize on different input types, and a learned router decides which expert handles each one.

The original intuition: if a model needs to handle both protein sequences and DNA sequences, you could train one large network on both — but you'd spend compute on protein-aware weights when processing DNA, and vice versa. MoE splits those responsibilities. Expert 1 handles protein-like inputs, Expert 2 handles DNA-like inputs, and the router learns when to call which.

In practice, modern MoE models — including Mixtral and reportedly GPT-4 — activate only a fraction of their total parameters on any given input. A model with 100 billion total parameters might behave like a 20 billion parameter model on any single forward pass. Faster, cheaper, and no loss of the representational power the full network provides — because each expert develops deep capability in its own domain rather than shallow capability across all domains. ProtLingo applies exactly this idea to protein biology, assigning different experts to different protein families.

Sign-off

That is THE AGENT SIGNAL for September 7th. Tomorrow we are watching for a formal Hormuz restricted-zone enforcement announcement — and whether any of the major agent framework teams pick up the Trace2Tower approach in their tooling. See you then.

Sources

1. AI’s Next Winners? Investor Bets on Snowflake, CrowdStrike and Palantir — Benzinga
2. Trace2Tower: Transition-Aware EigenTrace Induction of Multi-Level Skills for LLM Agents — arxiv.org
3. ProtLingo: Efficient Protein Language Modeling via Conditional Memory and Expert Routing — arxiv.org
4. Why are European countries moving their gold out of North America? — bbc.com
5. Iran to announce restricted zone outside Strait of Hormuz — aljazeera.com
6. At least 10 killed in central Mexico fireworks blast during festival — aljazeera.com
7. Alcaraz breezes into last eight at US Open — aljazeera.com
8. fastcore 2.2.22 — pypi.org