Findings from Morgan Stanley TMT 2026 and our predictions

All Newsletters12 March 2026

Prometheus is a figure from Greek mythology, he was a Titan who stole fire from the gods and gave it to humans. Zeus was enraged by this act, which he felt would embolden humans to stop fearing the God, so he punished Prometheus.

Prometheus syndrome is our Era’s version of the myth: where Society punishes fellow Humans for trying to steal intelligence from Humanity and give it to Machines.

It is not well known but the full title of Mary Shelley’s book was: “Frankenstein; or The modern Prometheus.”

Factsheet Commentary

For our full Factsheet, link here.

Morgan Stanley TMT Conference 2026

These are the key points that hit us hard from TMT.

As Jensen put it we are moving from a phase of the Era where Ai Answers to Ai Does. The Agentic wave is coming and its capabilities are revolutionary for enterprise workflows. Both OAI and ANTH spoke about how Agent Capability was the focus of their work for 2026. I would remind you that last year we said that the next step of progression that Ai needed to make in order to justify its ROIC was capable and credible functional agents. ANTH had told us they would be coming in Q4 25 to Q1 26, and wow have they delivered.

Personally, I can not express how I think this is a Paradigm Shift for Ai. For us, Ai has always been a substitute for the Economic Unit of Human Labour, just as Software 1.0 was. The Luddites belittle the predictions of these Tech pioneers but so far their predictions have be on point. So that is why point 2 below is such a heavy hitter for me.

Exponential Scaling. It gets better, as both AOI and ANTH said that the models would not just continue to scale in capability but the exponential was accelerating. Dario used the classic “Rice on a Chessboard” fable which illustrates exponential growth: 1 grain on the first square, 2 on the second, 4 on the third, and so on.

Squares 1–32: The growth is massive in absolute numbers, but manageable. The first half of the board presumably represents the journey from GPT-3 to the more sophisticated models we have today (like Claude 3.5 or Opus 4.6).
Square 40: Amodei’s point is that we have just crossed the midpoint. On the 40th square, the “doubling” effect means that the progress in this single step is greater than the entire sum of the previous 39 squares combined.
The “Second Half”: He argues that the next 24 squares (the remainder of the board) will move at a pace that “defies our imagination,” dismissing any talk of “scaling walls” or plateaus.

Amodei framed 2026 as the year of “Radical Acceleration.” He noted that the speed at which AI is currently being adopted in software engineering (coding) is the “canary in the coal mine”—a signal that this exponential surge is about to hit every other industry with the same force.

Jensen discussed about the old bragging metric for Enterprise size was human headcount and a correlation for Enterprise growth was growth in Seats/Employees. No longer. The WSJ ran an article whilst we were at TMT stating that CEOs with large Employee numbers were now the butt of jokes. The metric of the future is Token Use growth. By the way, how many employees do you employ and what’s your token spend pcm? I strongly suggest you should ask all companies you analyse what their monthly token spend is.

On a monthly add basis, Anthropic was adding ~$1B new ARR per month in late 2025 and just added ~$6B in ~19 days, which extrapolates to ~$10B a month, or 10x! This ARR supports the broader commentary on explosive token growth based on Claude Code, Agents, and OpenClaw.

This is literally mind blowing stuff for Technologists and the consequences for creation and disruption in the Economy & Financial Markets can not be underestimated. It is quite possible that from H1 2027 the USA starts to see a material productivity uplift from these agents. Note my prediction of this Productivity Shift starting 15 months away is because I believe the only remaining issue with the development of the Era of Ai is the speed of adoption by enterprises (thwarted by the incompetence of consultants) and government regulation. Sam called this the “capability overhang’ – the massive gap between what current AI models are technically capable of doing and how much of that power humans and businesses are actually using. However, agentic platforms like Claude Cowork are now easier than ever to adopt, handling much of the integration, connections and configuration themselves (our own Claude co-work even fixed an error in its own Config file!). Asimov may suggest that it may not be long before the Ai adopts ITSELF into the enterprise, removing the last human hurdle in the process.

My view is that the current productivity shift is being driven more by a reduction in bloated headcounts that were a remaining legacy of the Covid period.

In time, Productivity will rise driving further Economic Growth, and deflation not inflation, which will lead to spectacular Stock Market returns for the Ai Winners.

Encouragingly a number of the neo-clouds stated that they were seeing a growing demand from G2k Enterprise and Sovereign. Sam spoke very carefully about the clarity with which the Sovereigns see their position versus the Ai Labs and Sovereign opponents. I will not get drawn into this matter as it was discussed and predicted many years ago in depth by Isaac Asimov. Suffice to say that wise Sovereigns now see Ai Spend and Control as an existential question.

Jensen expounded on this point many times saying saying that the equation for the Enterprise in the future was simple. Compute => Intelligence => Revenue => GDP. Further to this it was often discussed by multiple parties that the correlation between Spend => Compute => Model capability was becoming a better understood correlation. NVDA’s message was BUY our AI factories or die. SNOW admitted that they built their own model, because they realised it would be existential if they didn’t.

What is also becoming clear is that Model capability and Compute capability are dynamic, not static so a version of a Model gets better with time modern AI systems scale their performance on the fly by utilizing test-time compute, reinforcement learning, and advanced reasoning to solve increasingly complex problems. The same is happening with Compute due to software optimisations so the abilities of say the H100 chip are now multi-fold when first released. This latter point, alongside the advancement of DC workflow Orchestration Software, allowing offloading of less complex inference to less complex chips, is extending the useful life and value of legacy GPUs. See chart below on H100 rental prices:

SAAS-pocalypse was discussed ad infinitum. We have discussed and predicted this for years so we are not going to go over old ground.

The best advice we did hear was from Hock Tan who to paraphrase: In Software you need to stick to the low levels of the Stack like the Operating System or just above. Application is risky and not sticky. Think of Ai as a piranha eating everything else. Engineers love using Ai, but hate using Software 1.0 which is restrictive. Can Ai replicate a whole ERP system, who knows in the future so why would I buy Software?

My view is IF you are asking which Software stocks to bottom fish you are depressingly missing the point! Software 1.0 is now Value investing whilst Ai investing is ultra-Growth investing. Software 1.0 needs a revolution (in Business Model) to survive, not evolution.

There was some discussion on other sectors and which are safe investments for the Ai future. My answer to this is always the same: The most valuable assets in the Future will be the same as the Past: Intelligence, Energy, Security & Entertainment. Another way to look at the same question is to ask which assets will not be deflated by the increased capacity of Ai? No that is not Software 1.0!

There was discussion at the conference about how Sports franchises were going to see strong growth as Humans are afforded increased leisure time which is exactly why we own FWONK US and TKO US. Remember, although the AIC has tried to rewrite our history and classify us as a Technology Fund, our prospectus clearly states we are a Global Growth Fund:

“The Company seeks to achieve its investment objective by constructing a diversified portfolio comprising of any of global equities and/or fixed interest securities and/or derivatives. The Company’s portfolio is focused on growth stocks. Growth stocks are those that are anticipated to grow at a rate above the average growth for the market.”

BTW just how well do you rate the AIC’s defence against SABA’s rampage through the sector? Someone quipped to me the other day that every Fund that SABA had attacked was a member of the AIC! Perhaps you may want to think about that correlation when you consider renewing!

So let’s move to Constraints. The DC spend is not going to be unlimited over the next 5 years because there are very clear constraints in a number of the required elements. Power is linear, whilst AI demand is exponential. In 3 years time, there simply will not be enough power to match the new DC demand. Many estimates revolve around 74GW of Power which must be brought onstream, we estimate 81GW:

We have taken two positions to play this secular theme. Below please find App Economy & Broker comments on both:

GE Vernova closed FY25 with a mic drop on future demand, reporting a staggering $22.2 billion in Q4 orders (+65% organic Y/Y), smashing analyst estimates of ~$18 billion. While revenue of $10.96 billion (+3.8% Y/Y) beat expectations, the story was entirely about the record-breaking backlog, which swelled by $15 billion sequentially to hit $150 billion.

The AI and data center boom is driving a supercycle for the Power and Electrification segments. Gas Power backlog and slot reservations jumped from 62 GW to 83 GW in a single quarter, and Electrification posted its largest order quarter in history. This momentum allowed management to raise FY26 guidance significantly: revenue is now expected to reach $44–$45 billion (up from $41–$42 billion).

Beth Kindig of the I/O fund has written an excellent Newsletter on Bloom Energy which can be found here: My Top 2026 Stock Pick for the AI Boom

There are other key points that come out of this sub point:

Constraints mean that you must maximise output per the constrained input hence Tokens per Watt $ is now the key parameter. This is a point BOTH Hock and Jensen brought up: when you are constrained then every Watt $ must be maximised so you must use the very best compute hence you will continue to choose NVDA/AVGO compute.
Several speakers (e.g. CRWV, INTC) said that NVDA is the most performant platform for compute. CRWV said that they build what their customers want and everyone just wants NVDA. Furiosa and D-matrix say they aren’t even trying to compete with NVDA, they are just trying to be a complementary chip for some workloads. OAI were highly complimentary too.
However, the constraints are not just in Power there are bottlenecks everywhere: GPUs, optics, power and data centre capacity are sold out well into ’27, which means capacity owners are gaining pricing power. Consensus capex estimates for 2026 now estimate +62% growth after being at just +18% 9 months ago. Every year this upward change in estimates happens and we doubt 2027 will be different. Many AI ecosystem players are now locking in supply of key Optical products and components with 2 – 3 year contract minimums. Remember the stocks we bought in December 2025, they were Optical plays.
The future will be multi-model orchestration across multi-compute architectures. We detailed the different compute types in the explanation of our latest NVDA model in our last Newsletter. It is noteworthy that LPUs can run at near 100% utilization, materially lowering energy per token vs. GPUs that idle 60-70% waiting on HBM. Hence, why NVDA acquired Groq.

The Vast TAM that Agents open up as a replacement for the Economic Unit of Human Labour.

Do you remember our analysis of the Jobs that can be replaced by Agents that we released in H1 2025?

Helpfully, ANTH have just released their own similar analysis:

So you may remember that OUR figures derived a 16% CAGR for Tech Spend:

Well ANTH’s figures would increase that CAGR to 22.7%!

Life Sciences => The Scientific Break throughs will start coming => this is how Ai sorts its PR problem.

Sam made a few very interesting model points. He pointed out that the prior belief was that models could never come up with new ideas as they are constrained to the known data they are trained on. Effectively, models were just ‘next token prediction’ machines. However, Sam is now confident that models have crossed the threshold to think and understand data and hence they are generating ideas which are ex-scope of their training data. On a more enterprise level, models will be a PROACTIVE system that watches your computer and your enterprise graph and workflows and then adapt to those in its suggested actions. He sees NO scaling wall and believes models released in the next few months/next few years will materially alter how businesses are managed.

This is a critical step because after the Agent Models take this next step in capability => the Ai native business will come on stream and start to outcompete the Legacy businesses. Hence the disruption will move from the Labour Unit level to the Enterprise level.

We look forward to Ai Native: Company Secretary services, Fund Accounting services, Payroll services, Share Registrar services etc. We will be early adopters.

Science/Medicine: Scientific breakthroughs are coming as OAI are giving scientists access to a very high compute model that thinks very hard about one problem so now once a week or every other week a scientist discovers that AI has solved their hardest problem. We see Healthcare / Life Sciences / Pharmaceuticals as areas ripe for AI-driven theoretical and prototype acceleration.

Vertical specific Industry Agents are coming, including Financials and Wealth Management – use them and win.

OpenAI Codex now >2M users growing >25% w/w which is unusual growth for a non-consumer products. Software engineering jobs have dramatically changed and expect this will happen to EVERY piece of Knowledge Work.

We heard from Perplexity who have developed orchestration models to assist particular Industry verticals including specifically financial services. It is suggested that you watch this very impressive video of their model creating a SpaceX DCF.

Amodei: Financial Services is ‘by far’ moving the fastest of all industries. ‘What’s happening in Coding will happen everywhere else, just more slowly .. Everyone else is on the same curve, just at a different point of that .. Models will be infused into every aspect of the economy’ first for writing code, but also using for processes / ‘scaffolding’ around that such as managing servers, controlling clusters, visually looking at features, software tools around it.

“Rule your mind or it will rule you.” Horace.

Cash is King and those with it can Fund the Future.

This point was made by Jensen which is unsurprising, but it was also made by others. We are entering a period of vast technological change that will require huge capital investment. The History of Technological change has predominantly been funded by the richest entities in the world, not by Banks.

The Age of Exploration was funded by the Royalty of Portugal, Spain, France & Britain. The Gilded Age including the Railways were funded by the richest industrialists like the Vanderbilts and Rockefellers. Richard Arkwright who built and funded the British Industrial Revolution ended up lending the Bank of England funds to execute the naval wars with France & Spain. So I don’t really understand why everyone is surprised that yet again the richest entities in the world are funding this Technological build out.

Ah you say but these always led to overbuild and waste – look how many of these railway lines are used today? That is a daft point as travel technology has just moved on. Let me present you a better example: Housing Technology. In London from 1680 to the 1900s there was a huge capital expenditure build out called stone or brick Town Housing with the Mayfair and Knightsbridge estates being built. Bubbles were proclaimed regularly but the Boom went on and on and on. As I wander home each night through these areas, I see very little signs of an overbuild!

I am always interested in the alternative future narrative of what the consensus is so certain of today. Are we sure that all these investments the likes of the Mag7 are making in the Era of Ai build out are so dumb? Either way several speakers at the conference said they saw an approaching period where capital became constrained and those with Cash would be very powerful playmakers.

Physical Ai.

Finally, Jensen kept reiterating a prediction which he has several times in the last 6 months.. May I remind you that Jensen has form for seeing the future and he claims that the next phase of the Era of Ai will be even bigger than the current DC build out. You can position now and ride that forthcoming wave, imagine if he is correct about its scale? The next wave is Physical Ai which means Robotics, Autonomous Vehicles etc. To be clear, we do not see this scaling materially for a couple of years but the opportunity is huge. The interesting issue with this prediction is that currently China looks to be materially ahead of the USA in Robotic Hardware technology.

For those interested in the further future, Richard also heard a presentation from a Neuromorphic computing start up arguing that Digital AI has a ceiling. We are running out of power so the only path forward if we want “Brain-level Intelligence” that doesn’t require its own dedicated power plant is Neuromorphic. Obviously, this sails against the fusion enthusiasts that claim that, in time, Fusion will provide limitless cheap power.

We are starting to digress…

Conclusions.

Last year, I predicted self-coding agents. This year we have frameworks (like DSPy) where models can systematically evaluate and rewrite their own internal instructions, system prompts, or “Chain of Thought” reasoning paths to get better results. They aren’t writing traditional code to change themselves, but they are coding their own tools and optimizing the natural language instructions that govern their behaviour.
Next year, we will have agents that are given read/write access to certain modular parts of their own architecture. So IF an agent realizes its memory retrieval (RAG) is inefficient, it might write a custom, highly optimized search algorithm, compile it, and permanently swap out its default RAG module for the new one. It is still bound by its core neural network, but it can upgrade its own peripheral modular or plug in software.
Last year, I predicted the Capex Boom would continue through 2025. This year, I predict the Capex Boom will continue to at least Mid 2027. This will be driven by exploding Token Use.
Last year, I hoped the functional agents were coming. Next year, you will see these agents being used by Ai Native start up that will compete with legacy service businesses in the Knowledge Economy. Adapt or Die.
Last year, I predicted that DC Workflow Orchestration Software would give legacy chips longer lifecycles. This year, we will see more orchestration of the multi-chip architecture DC with more specific purpose chips being developed for different models like World Models. Amodei stressed the importance of Anthropic’s “diversity of clouds (AMZN, GOOGL, MSFT) and chips (AWS Trainium, GOOGL TPU, NVDA GPU) as an advantage”.
Last year, we predicted the death of Software 1.0 Legacy stocks. This year, we do not predict a Lazarean Revival. This year, we predict we will see the death of some surprising legacy companies that are destroyed by fast adoption Ai Natives. One commentator quipped that the Sectors most ripe for disruption are those where the human workers wear suits but have limited credible academic or professional qualifications. Non graduate, Ex-High Street Bank clerks advising you on a Private Credit fund springs to mind.
Biggest Risk now: Humans especially Politicians. Large regulated industries and very large Enterprises in general have obstacles to overcome where you need distribution partners to help promote adoption. Giant behemoth’s have difficulty changing directions. Politicians will use Ludditism for their own political gains. Altman says he expected the three big ‘flash points’ with the Superintelligence take-off to be (1) are companies developing AI or the government more powerful? (2) how to reconfigure the economy when nobody can outwork a GPU? (3) when Superintelligence becomes powerful then who decides the values of alignment?

Amodei: “will see ‘exponential’ improvement” ‘It’s already big and going to get a million times bigger’.

I do not want to be melodramatic, but there is a point in the development of any technology when it breaks through an inflection point and to me this seems to be now.

“Thy Godlike crime was to be kind, To render with thy precepts less The sum of human wretchedness, And strengthen Man with his own mind.” — Lord Byron, from the poem Prometheus (1816).

NVIDIA visit.

New phase of Ai: Agentic, and its compute heavy. Things are tight. Have very good visibility for the year ahead. Focus was Supply. $100bn Purchase Commitments. Vera Rubin is coming. Samples have gone out. GTC coming for announcements.

They have OAI commitments in the new agreement. NVDA is their Training provider for 5GW and this includes Inferencing provider. And OAI will use all the cloud providers so that means all cloud providers will use NVDA. Same ANTH & META.

Inferencing costs need to come down with coding and agents. Token 10x per annum, Efficiency x5 so Chip Demand is 2x BUT actually its probably more. Because Token growth is coming. 2028 will be much different to today – massively different.

Investing in the Eco-system and need to help Ai players through the future. Next generation. A lot of cabling and connectors. Building relationships in Ecosystem is first Opportunity, BUT cash flow will be huge so they have room for BUY Backs too.

TSMC will give them the capacity to build what they need. They are very helpful.

When you go full DC scale system you get huge efficiencies. 9GW of Blackwell most is used at Inference. TCO, cost per token, cost per token watt. Cooling and ARM CPU were key to getting these costs down. They are no1 in Inference across world.

The Hyperscalers are always short so they are starting to think Longer Term. So expect Longer Term commitments => Valuation Uplift. Next 2 years will be VERY transformational.

Some quotes from Morgan Stanley:

“NVDA CEO described OpenClaw as ‘the single most important release of Software probably ever’ — it surpassed Linux as the most downloaded open-source software in just three weeks, while NVDA said Agents consume 1 million times more tokens.”

“MSFT contrasted GOOGL’s in-house TPU strategy and said ‘never get too high on your own supply’, meaning he wants heterogeneous software to run on the most cutting edge NVDA chips, in addition to in-house solutions.”

“NVDA neocloud clients CRWV & NBIS sound strong on demand outstripping supply.

Key Manchester & London Updates

Please remember that all Fund based news is now posted to: https://www.linkedin.com/company/mnl-ln

Twitter will only see posts on Ai, Technology and the global Economy.

We have released a new presentation for Q1 2026 which updates on our thinking for the year ahead. Please do review this presentation here.

Key Tweets of the Month

The Long Portfolio

We regularly post the Portfolio to this site: https://www.linkedin.com/company/mnl-ln please do click the Follow button as we may stop posting it here too in this section.