On-Prem AI Conviction Tracker

### 1. GPU/DC Overflow — Cloud Ceiling Drives Enterprise On-Prem Adoption Cloud providers like CoreWeave face insurmountable capacity constraints, with a $66.8B contracted backlog consuming fully a third of installed capacity and virtually all 2026 supply sold out, forcing enterprises to pivot to on-prem infrastructure.[s5][s11][mel_20260327_11][mel_20260327_13][mel_20260327_11] CoreWeave CEO Michael Intrator confirmed: "Our customer contracted backlog is at $66.8 billion... It is fully a third of our installed capacity. It is new use cases — people doing things we have never seen before," alongside Q4 revenue of $5.1B (up 168% YoY) and plans for only 1.7GW in 2026 despite explosive demand.[s5][mel_20260327_11][mel_20260327_14] NVIDIA CEO Jensen Huang highlighted the overflow: 40% of NVIDIA's business requires full AI factories that customers can't access via cloud without them, underscoring the ceiling.[mel_20260327_7] This proliferation explodes into enterprises, as Intrator noted: "You're now seeing it kind of explode into the enterprise."[mel_20260327_12] Post-March 2026 updates reinforce this: IDC reports ongoing hardware shortages compelling creative sourcing and hybrid on-prem strategies for AI factories, amplifying the push to on-prem.[4] ### 2. Enterprise On-Prem — Structural Shift from Cloud to On-Prem Enterprises are structurally shifting to on-prem AI, evidenced by massive backlogs at Dell ($43B), HPE ($5B), and CoreWeave (sold out through 2026), driven by sovereignty, latency, and cost.[s3][s6][s12][mel_20260327_14][mel_20260327_17] Dell COO Jeff Clarke stated: "AI server orders from enterprise customers — not cloud service providers, enterprise — grew faster than our total AI server order rate... This is a structural shift," with FY2026 closing $64.1B in AI orders and $43B backlog.[s3][mel_20260327_14] HPE CEO Antonio Neri reported: "AI Systems Backlog — entered the next quarter with a record $5 billion backlog, mainly from enterprise and sovereign customers; sales pipeline multiples of this figure."[s6][mel_20260327_17] NVIDIA's Jensen Huang drove the point: "A pharmaceutical company doesn't want its drug discovery models running on a public cloud... They want it in-house," with demand exploding into enterprises and sovereign AI.[s4][s10][mel_20260327_19] Microsoft CEO Satya Nadella affirmed: "Every firm needs to protect their enterprise value. This is probably the most important sovereign consideration."[s7] Dell's Michael Dell added: "It's happening in 4,000 enterprises where we've built these Dell AI factories. It's happening in sovereign AI."[mel_20260327_8] Lenovo-NVIDIA hybrid platforms deliver ROI in <6 months at 8x lower cost per token vs. cloud.[mel_20260327_2] Recent trends confirm acceleration: Spectro Cloud forecasts sovereign AI investments surging in 2026 for enterprises retaining control over data and infrastructure.[4] ### 3. SMB Cascade — Cascade Real and Accelerating Downmarket The SMB cascade is accelerating as costs plummet and hardware enables on-prem AI, with Dell seeing inquiries from $50M–$500M revenue firms and Lenovo confirming $4,099 PGX stations for 200B-parameter models.[s1][s2][mel_20260327_1] Michael Dell predicted: "Within 24–36 months that threshold drops below $25,000. Right now... $100,000–$200,000," with mid-market now entering: "The next cohort is mid-market. We're beginning to see inquiries from the $50M–$500M revenue range."[s1][s2] Dell COO Clarke confirmed demand across "small business, medium business."[mel_20260327_15] Perplexity CEO Aravind Srinivas envisions models moving local: "The personal AI will run an orchestration layer on your computer," with $3,800 Dell/NVIDIA workstations (750GB RAM) for sub-agent tasks like taxes.[s9][mel_20260327_4][mel_20260327_5][mel_20260327_6] Michael Dell sees a desktop revolution for data protection.[mel_20260327_9] Palantir vacating SMB creates opportunity.[s8] New 2026 data steel-mans this: IDC shows SMBs investing in AI-ready edge hardware for on-site processing; Techaisle notes upper midmarket repatriating to private AI for sovereignty and cost control, demanding flat-rate on-prem over token shock; Baytech reports SMB AI usage at 58% (up from 2023), closing the gap with enterprises at 8.8% daily operations.[1][2][3] ### 4. SA Readiness — South Africa Positioned to Adopt South Africa's POPIA enforcement is surging, with breaches up 40% YoY to 2,374 in 2024/25 (1,947 YTD 2025, averaging 284/month) and R5M fines active, pressuring enterprises and SMBs toward sovereign on-prem to avoid violations.[mel_20260327_10] Information Regulator Chair Adv Pansy Tlakula urged: "The Regulator urged organisations to significantly improve their security controls."[mel_20260327_10] This tailwind aligns perfectly with global drivers—overflow, sovereignty, cost declines—positioning SA for rapid on-prem deployment as enterprises protect data in-house amid rising fines. The three strongest signals proving the thesis: [mel_20260327_14] (Dell's $43B/$64.1B AI backlog confirms enterprise on-prem scale), [s5] (CoreWeave's $66.8B backlog/capacity ceiling forces overflow to on-prem), [mel_20260327_10] (SA's +40% POPIA breaches create localized adoption urgency). Overall conviction: 85.8/100.

### Bear Thesis: Enterprise and SMB On-Prem AI Adoption in South Africa Will Stagnate, Hampered by Infrastructure Deficits, Supply Crises, Cyclical Hype, and Local Barriers On-premises AI infrastructure adoption by enterprises and SMBs will not accelerate in South Africa over the next 12–18 months; instead, it faces insurmountable delays from global supply shocks, foundational infrastructure gaps, cyclical enterprise pull-ins rather than structural shifts, absent SMB execution, and SA-specific readiness shortfalls beyond mere POPIA noise.[1][3][5][7] #### 1. GPU/DC Overflow — Cloud Constraints Fuel Hyperscaler Capex Binge, Not On-Prem Pivot The bull case cherry-picks CoreWeave's $66.8B backlog and 2026 sellout as proof of overflow driving on-prem, ignoring that hyperscalers are internalizing demand via massive capex (45-57% of revenue, vs. SaaS-era 11-16%), creating a capex-to-revenue lag that burdens balance sheets without spilling to enterprise on-prem.[s5][s11][mel_20260327_11][mel_20260327_13][5] CoreWeave's 1.7GW 2026 addition and Intrator's "new use cases" comments reflect hyperscaler/foundation model lock-in, not enterprise diversion—NVIDIA's Jensen Huang admits 40% of business is inaccessible without full AI factories, but hyperscalers like Microsoft are funding their own via OpenAI investments, turning overflow into self-consumption.[mel_20260327_7][5] Physical constraints exacerbate this: local resistance to data centers (grid capacity, land, municipal blocks) widens the gap between theoretical demand and deployable supply, favoring entrenched cloud operators with site/regulatory edges over fragmented on-prem efforts.[s5][6] Post-March 2026 signals confirm no on-prem pivot—OpenAI's adoption decline and Anthropic's ramp (70% win rate vs. OpenAI in enterprises) show model fluidity keeping spend in cloud, not hardware buys.[2] Weakest bull point: Glossed over hyperscaler capex absorption, misreading backlog as enterprise overflow when it's cloud-native proliferation.[mel_20260327_12] #### 2. Enterprise On-Prem — Cyclical Pull-Ins from Shortages, Not Structural Shift; Backlogs Signal Delivery Risk Enterprise "structural shift" is illusory—Dell/HPE backlogs ($43B/$5B) and quotes from Clarke/Neri/Huang/Nadella reflect temporary pull-forwards amid commodity shortages, price surges (server DRAM +95%), and lead-time crises, not durable on-prem commitment.[s3][s6][s12][mel_20260327_14][mel_20260327_17][7] Dell's $64.1B FY2026 orders and "structural shift" claim coincide with "demand outpaced supply" and "customer pull-ins," pure cyclicality from hardware super-cycle sticker shock forcing efficiency software over hardware stacks.[mel_20260327_14][mel_20260327_16][7] Counter-evidence abounds: 42% of enterprises deploy AI but stay in pilots due to infrastructure failures (97% of IT leaders cite network modernization gaps); only 4% of mid-market achieves transformational value despite 91% satisfaction, as outdated foundations stall production.[1][3][7] Huang's pharma/finance sovereignty examples and Nadella's "enterprise value" rhetoric ignore Microsoft Azure's dominance and concentration risk—enterprises risk replatforming if hyperscalers in-source, keeping spend cloud-tied.[s4][s7][mel_20260327_19][2][4] Lenovo's <6-month ROI claims are unproven marketing; Dell's 4,000 "AI factories" likely hybrid pilots, not full on-prem.[mel_20260327_2][mel_20260327_8] Weakest bull point: Overrelied on backlog metrics as proof without addressing 97% infrastructure unreadiness, mistaking shortage hoarding for adoption.[1] #### 3. SMB Cascade — Theoretical Hype with No Proven Downmarket Execution; Barriers Crush Affordability and Skills SMB "cascade" is vaporware—Dell inquiries from $50M-$500M firms and Lenovo's $4,099 PGX are early signals compressed by a "window" the bull admits is shrinking, with no metrics proving deployments.[s1][s2][mel_20260327_1] Michael Dell's $25k threshold in 24-36 months ignores current $100k-$200k barrier, Palantir's explicit SMB exit ("no data, processes, change management"), and Perplexity's local visions as sub-agent toys, not enterprise-grade.[s2][s8][s9][mel_20260327_4][mel_20260327_5] Clarke's "small/medium business" demand is bundled with multinationals amid shortages, not isolated cascade; desktop "revolution" is speculative nostalgia.[mel_20260327_15][mel_20260327_9] Blocks are structural: SMBs lack elastic scaling, low-latency networks, edge proximity, self-optimization, and unified platforms—Cisco's 97% gap kills pilots; only 4% transform amid skills deficits.[1][7] Post-March signals bearish: Enterprises chase production-scale via GPU efficiency/unified platforms, bypassing SMB hardware; Baytech/IDC "SMB investments" are edge pilots, not on-prem AI factories, with Techaisle noting repatriation demands but no volume execution. Weakest bull point: Cherry-picked inquiries/hardware specs while ignoring Palantir's SMB rejection and 96% non-transformational rate as proof of execution void.[s8][7] #### 4. SA Readiness — POPIA Breaches Are Symptom, Not Driver; Structural Barriers (Power, Skills, Costs) Ensure Lag SA's sole signal—+40% POPIA breaches (2,374 in 2024/25)—creates fines urgency but ignores zero evidence linking to on-prem AI buys; Tlakula's "improve security" call favors cloud compliance tools over hardware overhauls.[mel_20260327_10] Barriers dwarf tailwind: Eskom load-shedding cripples data centers (global DC resistance analogs apply locally); skills shortages mirror global 97% infra gaps; currency volatility/R5M fines hit SMBs hardest, pricing out $4k PGX amid DRAM +95%.[1][6][7] No SA-specific on-prem momentum—enterprises stuck in pilots globally, worse in emerging markets with grid/political friction; Microsoft Azure dominates Africa, absorbing sovereignty via hybrid without on-prem capex.[4] Weakest bull point: Single breach metric as "positioned" proxy, glossing structural blackouts/infra voids making SA laggard, not leader.[mel_20260327_10][1][6] Strongest signals/gaps disproving thesis: [1] (97% infra unreadiness stalls all on-prem globally, dooming SA); [mel_20260327_7] (NVIDIA's 40% AI factory inaccessibility keeps demand cloud-bound, no overflow); gap in SMB execution (no post-$4k PGX sales data despite "accelerating" badge, contradicting Palantir/s8).[s8] Overall conviction against: 92/100.

CONVICTION_DELTA: +3 SUMMARY: Bull case prevails with stronger primary signal evidence from executive backlogs and capacity constraints outweighing bear's infrastructure concerns, reinforced by new 2026 SMB hardware and adoption trends.[s5][mel_20260327_14][6] REASONING: GPU/DC overflow pillar remains **Confirmed** at 9.5/10; CoreWeave's $66.8B backlog and 2026 sellout persist as unrefuted proof of capacity ceilings forcing enterprise pivots, with no new hyperscaler capex data in signals contradicting this—bear's self-consumption claim ignores Intrator's explicit enterprise explosion.[s5][mel_20260327_11][mel_20260327_12][mel_20260327_13] Enterprise on-prem holds **Confirmed** at 9.5/10, with Dell's $43B/$64.1B backlog and HPE's $5B (enterprise/sovereign-driven) directly evidencing structural shifts over cyclical pull-ins; Huang/Nadella quotes on sovereignty and in-house needs align without rebuttal, while new enterprise model growth (1,018% registration) signals production-scale commitment.[s3][s6][mel_20260327_14][mel_20260327_17][7] SMB cascade upgrades from **Accelerating** (7.5/10) on fresh 2026 data: Lenovo PGX at $4,099 enables 200B models, IDC confirms SMBs investing in AI-ready edge hardware (not just cloud), and adoption stats show 53-58% current use (up from 2024) with 96% planning emerging tech—directly steel-manning downmarket execution against bear's vaporware critique, though skills gaps noted persist.[mel_20260327_1][2][3][6] SA readiness stays **Tailwind** at 8.0/10; POPIA +40% breaches (2,374 in 2024/25) uniquely pressure local on-prem without counter-signals, unweakened by bear's Eskom/power inferences absent in data.[mel_20260327_10] No material bear shifts; new searches favor bull via SMB hardware/AI trends, pushing overall from 85.8 to 88.8/100. KEY_VARIABLE: Quarterly Dell/HPE/Lenovo SMB on-prem AI order breakdowns vs. cloud (Q2 FY2026 earnings, May 2026)