Skip to content
DEEP VERTICALALPHA

LATEST ESSAY / 2 OCTOBER 2026

There is no
memory cycle.

AI work can pause while useful state remains. What follows for the machines that make intelligence possible?

A new essay connects model memory, agent state and physical infrastructure. The argument challenges a familiar commodity-cycle template; prices and margins still need evidence.

JOHN COLLINS · DEEP VERTICAL ALPHASELECTED RESEARCH, SHARED.
WORK CAN PAUSE. STATE CAN PERSIST.CONCEPTUAL MODEL
EXECUTION ACTIVEMODEL WEIGHTSLearned parametersSTORE & SERVECONTEXT CACHETemporary working contextRETAIN OR RECOMPUTEWORKSPACETask state between actionsKEEP WHILE WAITINGDURABLE RECORDSelected lessons for laterSELECT & RETRIEVESTATE RETAINED SELECTIVELYEXECUTION ACTIVEWeightsLEARNEDContextTEMPORARYWorkspaceTASK STATERecordSELECTED LESSONSDIFFERENT ROLES · SELECTIVE RETENTION

Execution draws on different kinds of state.

Weights, context, workspace and lasting records are not interchangeable. Their placement shapes capacity, bandwidth and latency.

A conceptual explanation, not measured utilisation or a demand forecast. Retention and placement vary by architecture.

A RESEARCH NOTEBOOK, SHARED

Look closely. Trace the relationship. Challenge the explanation.

Join the discussion

01 / SELECTED RESEARCH

The interesting part
is the missing link.

Public issuer evidence, source reconstruction and research methods. Specific studies that make a wider discipline tangible.

02 / THINK WITH THE WORK

Inspect the reasoning

Follow the cash.
Test the inference.

Does cash received establish surplus for shareholders?

Oracle · Q1 FY27 · USD billionsThree months ended 31 August 2026
01 / ReceiptObserved

The cash is real. Its source matters.

Observed
Oracle reported $23.103bn of operating cash flow. This already includes the $11.363bn specified increase in deferred revenue from customer prepayments with a significant financing component.
Inference
The advance provides genuine operating funding. Receiving it does not, by itself, establish earned surplus.
Unknown
How much of the associated work has been delivered and accepted, and at what remaining cost?

Cash flows, p. 5 · Customer prepayments, p. 7

One assumption, held apart

What changes if that inflow is excluded?

Remove only the specified prepayment inflow. Hold every other reported number fixed.

$0.000bn
Reported OCF
23.103
Excluded inflow
0.000
Reported capex
28.499
Arithmetic residual

Equals issuer-defined free cash flow at zero exclusion.

−5.396bn
-15-10-50
Issuer-defined FCF −5.396bnSelected residual

Not normalized or counterfactual cash flow, a valuation, or a forecast. This does not remove all advances or other working-capital movements. Equity financing is excluded from both calculations.

One issuer, one quarter. No new directional security judgment.

Inputs + formulas

03 / A WORKING DISCIPLINE

Useful insight.
Then the machinery.

Research becomes useful when a relationship becomes clearer. A source, a diagram and an experiment can reveal different parts of the same question. Each should make its assumptions inspectable.

Read the method & its limits
  1. 01

    Start with the evidence.

    Find the primary source, its date and its scope. An announcement, a deployment and a measured result are different facts.

  2. 02

    Follow the mechanism.

    Connect capability to constraints and consequences. Show which relationships are observed and which remain hypotheses.

  3. 03

    Try to break the inference.

    Test timing, comparison groups, magnitude and rival explanations. Keep the unsupported step open to challenge.

  4. 04

    Make the reasoning visible.

    Share sources, calculations and useful visual models. Label the assumptions and the limits of each example.

  5. 05

    Leave room for correction.

    Date the work, preserve its context and acknowledge what new evidence changes.

04 / RESEARCH IN THE OPEN

Share the evidence.
Keep the question open.

Deep Vertical Alpha is the public research notebook. Studies are selected for what they explain; they are dated so readers can distinguish the evidence then from what may be known now.

SCOPE & LIMITS

Examples draw on public sources. An issuer study does not imply an investment holding. A conceptual diagram explains a relationship; a calculation tests stated assumptions. Neither is a forecast.

How to read the work

ORACLE STUDY / EDITION 2026.39 · Evidence cutoff 15 Sep 2026, 17:35 SGT · Public presentation updated 2 Oct 2026

John Collins, PhDJOHN COLLINS / SINGAPORE

05 / THE RESEARCHER

John Collins.

A researcher’s curiosity.
A builder’s attention to how things work.

I investigate artificial intelligence and superintelligence: what new systems can do, what constrains them, and how their effects travel through technology and the wider economy.

I share selected findings here as Deep Vertical Alpha. Deep Vertical AI is the Singapore company through which I develop this research.

Earlier work spans investment-bank quantitative research, HSBC’s Artificial Intelligence and Data Analytics laboratory, FTI Consulting’s Asia AI practice, and doctoral research. The current work is here to inspect.