Uncategorized

SaaS Companies Face Extinction

Dark shadows are looming behind SaaS companies – Is this a bubble ?

Big software companies (Salesforce, ServiceNow, Workday, Adobe, etc.) need to add billions of incremental revenue to avoid decimation by GPU economics and investor expectations:

Part 1: The Core Problem – SaaS Can’t afford AI and be profitable?

AI workloads aren’t free. Whether companies buy GPUs (with rapid depreciation) or rent them in the cloud (paying hyperscaler margins), costs scale faster than SaaS pricing models allow. If margins fall, valuations collapse, because the market assumes SaaS margins remain flat or expand. To sustain valuations, AI usage must be monetised with software-like margins, not absorbed as infrastructure costs.

Without a viable downstream pricing model, infrastructure players (NVIDIA, CoreWeave, etc.) can’t support sustainable growth, and the current buildout risks becoming a bubble.

---

Part 2: Token Use by User Type

User Type	Monthly Tokens (MT)
Classic SaaS seat (light AI, speech)	0.6–1.5
Enhanced SaaS seat (heavy copilots, call center)	3–6
Palantir / advanced analyst	15–30
Medical imaging (radiology)	150–600
Media analysis (surveillance, video)	1,500–3,000
Media generation (video/VFX)	3,000–30,000+

---

Part 2a: Imputed GPU Costs (Blackwell at ~$1.26/MT)

User Type	Monthly Use (MT)	Blackwell Cost / Month
Classic SaaS seat	0.6–1.5	$0.75–$1.90
Enhanced SaaS seat	3–6	$3.80–$7.60
Palantir / analyst seat	15–30	$19–$38
Medical imaging seat	150–600	$190–$760
Media analysis seat	1,500–3,000	$1,900–$3,800
Media generation seat	3,000–30,000+	$3,800–$38,000+

---

Part 3: Examples

Let’s model costs and margin to work out the revenue that is needed to continue

• Classic SaaS (Salesforce/ServiceNow): small AI uplift per seat (summarisation, embeddings).
• Enhanced SaaS: reasoning agents, call center copilots.
• Palantir / analyst seat: orchestrated reasoning, multi-model queries.
• Medical imaging: each X-ray/CT study = tens of thousands of image tokens.
• Media analysis: surveillance or legal video review, millions of tokens per hour.
• Media generation: VFX or AI video, billions of tokens per project.

---

Part 4: Financial Lens

Additional Pressure on Top 10 Software Companies

The top ten Nasdaq-listed software companies together generate more than $350B in annual revenue. To maintain ~20% EBIT growth while covering AI compute, they will collectively need to add on the order of $70B+ new revenue each year. This is before accounting for incremental GPU costs, which will further raise the bar. AI seat pricing therefore becomes central not only to product strategy but also to market valuation.

These figures are based on current user numbers and revenue baselines assume similar number of seats (not growth of users). If this revenue is from more users, the required revenue growth would be even higher to preserve margins.

Breakdown of approximate revenue growth needed (20% p.a.):

Company	Current Revenue	Est. Users	Growth Needed ($B)	Growth/User/Year
Microsoft (software segment)	~$85B	~400M	~$17B	~$42
Oracle	~$50B	~50M	~$10B	~$200
Adobe	~$20B	~30M	~$4B	~$133
Salesforce	~$35B	~20M	~$7B	~$350
Intuit	~$16B	~100M	~$3B	~$30
ServiceNow	~$10B	~1M	~$2B	~$2,000
Workday	~$7B	~60M	~$1.4B	~$23
Autodesk	~$6B	~6M	~$1.2B	~$200
Atlassian	~$4B	~20M	~$0.8B	~$40
Snowflake	~$3B	~1.5M	~$0.6B	~$400

Together this aligns with ~$70B+ revenue growth per year required just to keep EBIT compounding at 20%.

Hardware Efficiency vs. Economics

GPU Model	Approx Cost / MT
A100	~$3.82
H100	~$3.04
Blackwell	~$1.26

Blackwell GPUs reduce cost per MegaToken significantly compared to A100 and H100, often by more than half. But even at ~$1.26/MT, the scale of token usage in advanced workloads means absolute costs remain material. For SaaS companies, efficiency gains ease the curve but do not remove the structural need to recover AI expenses through pricing.

Vulnerability of Infrastructure Providers

This dynamic also exposes risk for NVIDIA and cloud infrastructure companies. Once the upfront scale-out of GPUs is complete, if SaaS vendors fail to pass costs downstream, usage growth does not automatically translate into proportional revenue growth for infrastructure. Margins for the supply side could contract sharply, making the current buildout fragile if the economics of AI seats do not align with sustainable pricing. In other words, NVIDIA and infrastructure providers are highly exposed: once the upfront scale-out wave passes, their long-term growth depends on whether software companies can successfully monetise AI tokens without eroding their own valuations.

The market is surprisingly consistent in how it values hypergrowth tech.
No matter how much hype surrounds a company, the multiples compress in a predictable pattern as scale increases.

Story → Scaling → Sustained

Story

• Narrative and TAM drive valuations, not fundamentals.
• Growth is 2–3× YoY.
• Market assigns 35–55× ARR.
• Examples: OpenAI in 2023–24 (tiny revenue, $300B valuation), Snowflake at IPO in 2020.

Scaling

• Execution and enterprise adoption become the focus.
• Growth slows to 50–100% YoY, but ARR climbs into the billions.
• Multiples compress to 20–30×.
• Examples: Anthropic today (2025, $5–10B ARR), Datadog during 2018–21.

Sustained

• Market leadership is entrenched, with durable ~30% growth even at scale.
• Multiples compress further to 10–20×, but on much larger revenue bases, driving multi-trillion valuations.
• Examples: NVIDIA in the 2020s (>$60B revenue, still 30–40% growth), AWS in the late 2010s ($40–60B revenue, ~30% growth), Google Ads at $100B+ scale.

---

Anthropic Today

• Transitioning from Scaling toward Sustained.
• ARR trajectory:
  • ~$1B (Jan 2025)
  • ~$3B (May 2025)
  • $5B (Aug 2025)
  • Projected $7–8B exit 2025 organically.
• With a Microsoft Copilot allocation (for example a SWAG of 25% of MS’s AI vendor budget): projected $10B exit ARR in 2025.

---

Market Multiples

• OpenAI: valued at ~$300B on ~$5.5B ARR (~55×).
• Anthropic: valued at ~$183B on ~$5B ARR (~36×).
• Historical glide path:
  • Story = 35–55× ARR
  • Scaling = 20–30× ARR
  • Sustained = 10–20× revenue

---

Anthropic Valuation Model (with & without Microsoft)

Year	ARR (organic)	ARR (with MS)	Multiple (range)	Valuation Band
2025	~$7.5B	~$10B	~30×	$225–300B
2026	~$13.5B	~$16B	25–28×	$338–448B
2027	~$20B	~$24B	22–25×	$440–600B

---

Takeaway: Anthropic is leaving the Story stage behind and entering Scaling → Sustained. Even with multiple compression baked in, it scales from ~$200B today toward $0.5–0.6T by 2027, with Microsoft providing the kicker that accelerates the transition into the Sustained phase.

On 24 February 2022, Russian troops surged out of Belarus toward Kyiv, barely 150 km away. The Kremlin expected the capital to fall in days.

A snail, crawling nonstop, could cover that distance in about five years. More than three and a half years later, Russia still hasn’t managed it. Kyiv endures, the frontline frozen.

The image is almost absurd: an army of tanks and missiles, outpaced by the patience of a garden snail. But it’s also bittersweet. The failure reflects not only Russian weakness, but the staggering cost Ukraine has paid to hold its ground.

Broadly Intelligent AI: Why Hardware Matters as Much as Algorithms

Artificial Intelligence has surged forward, but the next phase is less about clever algorithms and more about the hardware that makes true intelligence possible. To move from narrow, task-specific models to Broadly Intelligent AI, the industry must solve a scaling challenge that is both technical and economic.

How AI Inference Works Today

Every AI interaction follows the same loop:

• Inputs → tokens — Speech, documents, and images are broken into tokens — numerical fragments a model can process. This step already uses GPU compute.
• GPU inference — Tokens are run through the model on GPUs, where billions of parallel calculations predict the next token.
• Context window — Tokens are temporarily stored in a context window — the model’s short-term working memory.
• GPU output — The model’s prediction is converted back into text, responses, or structured outputs. CPUs format results, but GPUs provide the intelligence.

The Context Window Problem

• Current limit: Models handle 4k–32k tokens — a few pages of text or a short conversation.
• Need for Broad AI: Broadly Intelligent AI requires millions of tokens — the ability to reason across books, sustained conversations, or months of data.
• The failure mode: Once the window fills, older tokens are dropped. The AI forgets, preventing broad reasoning across long timelines.

Scaling Is Exponential

Expanding the context window isn’t linear — it’s exponential:

• Attention compares every token with every other → O(n²) scaling.
• 10× bigger context → 100× more compute.
• 1,000× bigger context → 1,000,000× more compute.

This strains every part of the stack:

• Memory: from GBs today → TBs tomorrow.
• Bandwidth: from ~1 TB/s → multi-TB/s.
• Compute: requires thousands of times more parallelism.
• Interconnects: GPUs must share data at unprecedented speed.

Possible Compromises

Researchers are testing fuzzy memory architectures:

• Recent tokens stored with precision.
• Older tokens compressed or approximated.

Benefit: reduced compute and memory cost.
Risk: weaker accuracy, introducing human-like “misremembering.”
Verdict: more research needed before this underpins Broad AI.

Market Landscape

• NVIDIA: Iterative GPU scaling, faster interconnects, denser memory.
• Cerebras: Wafer-scale processors with massive on-chip compute and memory.
• Startups: Groq, Tenstorrent, Graphcore — each targeting efficiency and latency.
• Memory innovators: HBM, stacked DRAM, photonics — essential to breaking bandwidth ceilings.

Investor Takeaways

• Context scaling is the bottleneck — it is the gating factor to Broad AI.
• Exponential scaling curve — costs grow faster than most forecasts assume.
• Winners bend the curve — whoever delivers exponential capacity without exponential cost will dominate.
• Opportunities extend beyond models — hardware, semiconductors, and memory will define the landscape as much as software.

Looking Forward

The road to Broadly Intelligent AI demands:

• Larger, affordable context windows.
• Smarter memory architectures balancing accuracy and efficiency.
• Hardware leaps — wafer-scale, photonics, stacked memory — not just incremental GPU upgrades.

The open question: will NVIDIA’s scaling, Cerebras’ radical design, or a yet-unseen breakthrough define the future? What is certain is that Broadly Intelligent AI will not emerge from software alone. Hardware evolution is equally critical.

Palantir continues to post standout numbers in 2025. Revenue is on track to grow ~45% year-over-year, with U.S. commercial sales nearly doubling and government contracts expanding more than 50%. Profitability is improving too, with net margins already at 33% and expected to expand steadily.

But valuation tells a different story. At today’s levels, Palantir trades at one of the highest revenue multiples in software. If that premium compresses—as history suggests it will—the company’s financial gains may not translate into major upside for investors.

---

Projection With Multiple Compression

Assumptions:

• Revenue growth: +45% per year (2025–2028)
• P/S multiple: 20% compression per year (from ~74× today)
• Net margin: +1% per year (from 33% in 2025)

Year	Revenue ($B)	P/S Multiple	Market Cap ($B)	Net Margin	Net Income ($B)
2026	6.02	59.2	356	34%	2.05
2027	8.73	47.4	413	35%	3.05
2028	12.65	37.9	479	36%	4.55

---

What It Shows

• Revenue soars: more than triples in three years.
• Profitability scales: net income more than doubles, reaching $4.5B by 2028.
• Valuation growth slows: despite explosive fundamentals, market cap rises only modestly—from ~$370B today to ~$479B in 2028.

---

Why Upside May Be Limited

• Current multiples are historically elevated; even world-class growth cannot fully offset compression.
• Market cap growth (~+29% over three years) lags far behind revenue growth (~+200%).
• In effect, investors may see strong fundamentals but not commensurate share price appreciation.

---

Conclusion

Palantir is executing exceptionally well—fast growth, rising margins, robust cash generation. But the valuation already prices in much of that success. If multiples continue to normalize, the stock may deliver less upside than the business results suggest.

Anthropic is preparing to close a new funding round at a $170 billion valuation, a figure that would cement it as the second-most valuable private AI company behind OpenAI. But as significant as that milestone is, history suggests the real valuation story unfolds immediately after—in the secondary market.

If recent parallels from OpenAI and CoreWeave are any guide, Anthropic’s post-round pricing may rise sharply as investors who missed primary allocations rush to gain exposure. And that could push the true implied valuation well beyond $170B, at least in the near term.

What We’ve Seen Before

OpenAI

In March 2025, OpenAI raised at a ~$300B valuation via a staff-led secondary offering. Just months later, secondary bids surged toward $500B—a 67% jump—as institutional buyers aggressively sought access to the market leader. Much of the surge came from parties locked out of formal allocation channels, including sovereign wealth funds and crossover hedge funds.

CoreWeave

Pre-IPO secondaries priced CoreWeave near $23B. At IPO (March 2025), it listed at $27B, with public markets immediately absorbing the difference. The modest but clear +17% valuation uplift shows even infrastructure players benefit from liquidity timing.

Anthropic: Next in Line

Now Anthropic is on the verge of closing its own step-change round, jumping from its last $61.5B valuation to a proposed $170B. That’s nearly a 3x markup, and the market’s response is expected to be just as aggressive—if not more.

Why?

• Unfilled demand from Series E: Last time, secondary prices rose up to 25% above issue price within days.
• Clear market position: Unlike OpenAI, Anthropic offers enterprise buyers and investors a more neutral, governance-stable option.
• Strategic scarcity: With so few scaled AI platforms left in private hands, Anthropic is likely to be heavily oversubscribed again.

What We Expect

• Secondary pricing could land closer to $185–200B in the immediate aftermath
• Existing shareholders (Series C–E) could see liquidity offers well above their entry
• New allocators may prioritize speed over price to secure a seat

Comparative Summary

Company	Last Valuation	Post-Round Value	Premium
OpenAI	$300B	~$500B	+67%
CoreWeave	~$23B	$27B (IPO)	+17%
Anthropic	$170B (expected)	$185–200B (projected)	+9–18%

Conclusion

Anthropic’s $170B valuation won’t be the final word. In this market, oversubscribed rounds don’t end the conversation—they just shift it to secondary desks. And with investors already watching OpenAI soar to half a trillion and CoreWeave delivering a clean IPO premium, Anthropic could emerge as the next must-own AI name—at any price.

The question isn’t whether it hits $170B. It’s how fast it clears $200B once the ink is dry.

---

Frankly, now I don’t automatically believe AI’s answers anymore…

Has GPT-5 Already Hit Its IQ Ceiling?

TL;DR: I asked GPT-5 to create some IQ-130 questions. It came back with three that all failed the most basic test of validity — one had two correct answers, one was impossible to solve, and was unsolvable without another, hidden rule. If this is the best version of ChatGPT, its reliable reasoning may top out around an IQ-130 equivalent — strong for everyday use, but far from the flawless logic people imagine.

---

I Wasn’t Testing GPT-5 — At First

This didn’t begin as an experiment. I just wanted a few IQ tests designed to show my wife difficulty levels . I asked GPT-5 to produce 130 and 150 IQ level questions -to demonstrate the idea of IQ tests Instead I got something that made me pause — and then dig deeper.

---

What I Asked vs. What I Got

The questions arrived . At first glance, they looked legitimate. But when I tried solving them, they started to unravel.

---

Puzzle 1 – Did not have a unique answer

GPT-5’s question:
Five people—A, B, C, D, E—are seated in a row.

• A is to the left of C but not directly next to C.
• B is between D and C.
• E is at one end.
  Who is in the middle seat?

What I found: There are two valid arrangements: ChatGPT5 agreed

• A D B C E → middle is B
• E A D B C → middle is D

Two answers means it fails the one-solution rule that IQ tests demand.

---

I gave ChatGPT a chance to rework the question – It failed again and was Impossible to Solve

What I asked: “Now rewrite it so there’s only one possible answer.”

GPT-5’s question:
Five people—A, B, C, D, E—are seated in a row.

• E is at one end.
• A is to the left of C, exactly two seats away.
• B is immediately between D and C.
• D is not at either end.

What I found: The conditions contradict each other. There’s no seating arrangement that satisfies them all. Yet GPT-5 presented it as a solvable puzzle.

---

I then requested an IQ 150 level set of 3 questions – One failed

GPT-5’s question: A sequence that looked fine but required an unstated extra rule to get to GPT-5’s answer. Without guessing that hidden step, multiple answers fit — a direct breach of the principle that IQ tests must be fully self-contained.

---

The Pattern

In all three cases, GPT-5 produced puzzles that looked correct but failed under close scrutiny.

• Ambiguity: More than one answer fits.
• Contradiction: No answer exists.
• Omission: An essential rule is missing.

These are exactly the flaws that a competent human puzzle-setter — even at an average IQ — would usually avoid. Yet GPT-5 missed them.

---

What This Suggests About GPT-5’s “IQ”

This isn’t about trivia or specialist knowledge — these were reasoning problems. GPT-5 handled simple logic fine, but above a certain complexity, the cracks showed.

The emerging picture is a reliable reasoning ceiling around IQ 130 equivalent.

• At this level: GPT-5 can manage most everyday reasoning, pattern spotting, and problem-solving.
• Beyond this: The risk of hidden flaws, contradictions, or overconfident wrong answers rises sharply.

---

Why This Matters

If GPT-5 is the “best” version yet and still struggles beyond IQ-130-equivalent reasoning, that matters because:

• In research: Subtle flaws can mislead entire projects.
• In law, engineering, strategy: Confidently wrong answers can be costly.
• In public trust: People assume ChatGPT “must be right” because it sounds right.

The problem isn’t just that GPT-5 makes mistakes — humans do too. It’s that it doesn’t recognise when a problem is flawed. It will confidently produce an answer to a question that can’t be answered at all.

---

Bottom line: I didn’t set out to measure GPT-5’s IQ. But after three broken puzzles in a row, I’m left with a question: If it struggles above IQ 130, how much can we really trust it for deep, critical thinking?

---

A Bottom-Up, Valuation Framework For 2027

AI is powering a global economic transformation. By combining a bottom-up revenue model with valuation multiples proven by past hyper-growth companies like Salesforce and ServiceNow, we can make a disciplined and credible case for the near term value of companies like OpenAI, Anthropic, and NVIDIA.

---

Step 1: Bottom-Up AI Market Sizing

User Base Assumption:

• 4 billion people globally are touched by AI (50% of global population)

Average Monthly Spend per User:

• $5 for personal AI
• $20 for business use
• $4 for automation (e.g. AI cars, smart devices)
• Total: $29/month

Annualized Total Market:
$29 × 4B users × 12 months = $1.392 trillion

---

Step 2: AI Spend Breakdown by Layer

• Infrastructure: $696B (50%)
• Applications: $511.5B (36.7%)
• Core Software: $184.5B (13.3%)

---

Step 3: LLM Platform Revenue

Assuming 70% of Core Software spend goes to large language model (LLM) platforms:

• Total LLM spend: $129.2B

LLM Vendor Share:

• OpenAI (40%): $51.7B
• Anthropic (30%): $38.8B
• Others (30%): $38.8B

---

Step 4: NVIDIA’s AI Revenue

Assuming NVIDIA captures 60% of infrastructure spend:

• 60% of $696B = $417.6B

---

Step 5: Apply a Proven 14× Revenue Multiple

Why 14×?
During their breakout phases, companies like Salesforce, ServiceNow, and Shopify all traded at 10× to 20× revenue.
14× is a conservative, historically proven mid-range multiple during 50%+ growth phases.

---

Final Valuation Summary (14× Revenue)

OpenAI

• Revenue: $51.7B
• Valuation: $724B

Anthropic

• Revenue: $38.8B
• Valuation: $543B

NVIDIA

• Revenue: $417.6B
• Valuation: $5.85T

---

Why This Works

• Built from the bottom up — based on real users and real spending behavior
• Uses valuation multiples proven by high-growth tech companies
• Treats NVIDIA not just as a chip company, but as a software-style AI platform with deep infrastructure lock-in

---

This model doesn’t rely on inflated TAM or hype. It uses usage, spend, and precedent — and it suggests the next $10 trillion in enterprise value is already forming.

---

---

As we move to federated AI systems, those where continuous learning happens at the edge and fed back to the model then it’s essential to run on the same stack. For example if you want one robot to learn from another then this is a federated learning environment. The effective training at the edge has to be consistent with the training at the core and for most people that means NVIDIA CUDA.

In short,the next generation of AI isn’t just running in the cloud — it’s evolving in real time, at the edge. From smart cameras and autonomous machines to mobile devices and industrial sensors, edge systems are now being asked to reason, personalize, and learn locally. This shift blurs the line between training and inference, introducing new technical and operational demands.

NVIDIA, already the undisputed leader in AI training, is extending that dominance to the edge. Its CUDA ecosystem — encompassing GPUs, toolkits (like TensorRT, DeepStream, and Triton), and the Jetson platform — allows developers to train, optimize, deploy, and adapt models using a consistent, end-to-end architecture.

Crucially, inference is no longer a separate phase from training. In modern AI workflows — especially those involving continual learning, federated training, or adaptive models that blur training and inference —

Even in old, single shot AI like chat GPT 3 or 4 Inference was an iterative part of the training process It’s an integral part of the training loop.

When you then run inference on a different hardware or software stack, you’re introducing unpredictable behavior that isn’t accounted for during training, and that breaks the very assumptions the model was tested and validated on.

This mismatch creates real risk. Variations in numerical precision, operator behavior, quantization methods, and data handling can lead to silent accuracy degradation, failed model updates, and long-tail debugging costs.

Even more critically, the rise of reasoning models — which loop through multiple inference steps, reference memory, or adapt on the fly — depends on training-grade architectural fidelity. These models often behave more like learners than static classifiers, and they break down quickly when exported to stripped-down or non-aligned runtimes.

While edge hardware from other vendors (AMD, Intel, Apple, etc.) can run inference, they often lack the deep software alignment, full training support, and operator-level consistency required for advanced AI workflows. In contrast, NVIDIA offers a vertically integrated solution that reduces integration risk, improves model reliability, and accelerates deployment — all while enabling forward compatibility with evolving AI capabilities.

---

10 Reasons a Split Stack (e.g., Training on NVIDIA, Edge on AMD/Intel) Is Becoming Commercially Impractical

1. Accuracy suffers — different hardware can distort model behavior in unpredictable ways.
2. Harder to test and validate — verification breaks down when core and edge don’t match.
3. Learning feedback loops fail — federated learning and co-evolution rely on shared assumptions.
4. Debugging becomes expensive — inconsistent behaviors are harder to trace and fix.
5. Longer time-to-market — more QA cycles, more rework.
6. Higher operational risk — undetected errors at the edge can have outsized consequences.
7. Engineering burden rises — fragmented tools mean higher maintenance and lower velocity.
8. Advanced AI is harder to deploy — reasoning models and on-device learning need training-grade compatibility.
9. Cost of ownership increases — upfront savings are erased by long-term integration and support costs.
10. Strategically misaligned — the AI ecosystem is converging on unified, end-to-end platforms. Split stacks will fall behind.

---

Let me know if you’d like this formatted for a deck, summary slide, or adapted into a talking brief for investors or internal presentations.