Apple’s $599 MacBook Neo Beats a 16-Core Cloud Server — On Cold Queries
Apple’s $599 MacBook Neo Beats a 16-Core Cloud Server — On Cold QueriesDuckDB ran 43 analytical queries on Apple’s cheapest laptop. The results are… complicated.
0.57 seconds median cold query. 59.73 seconds total. $599 laptop vs. $0.768/hr cloud box with 4x the RAM. But wait — the hot-run numbers tell a different story entirely.
DuckDB’s team bought Apple’s brand-new MacBook Neo on launch day and pitted it against AWS instances costing thousands per month. The headlines write themselves: “budget laptop destroys the cloud.” The data says something more nuanced.
🧩 Dumb Mode Dictionary
| Term | What It Actually Means |
|---|---|
| DuckDB | An in-process analytical database. Think SQLite but for crunching data instead of storing app state |
| ClickBench | A benchmark with 43 queries that hammer aggregation and filtering on 100 million rows |
| TPC-DS | A harder benchmark — 24 tables, 99 queries, window functions, the works |
| Cold run | First query execution when nothing is cached. Disk speed dominates |
| Hot run | Repeat execution when the OS has cached data in memory. CPU dominates |
| NVMe SSD | Fast local storage soldered to the laptop. No network hop required |
| A18 Pro | Apple’s iPhone 16 Pro chip, now inside a $599 laptop. 2 fast cores + 4 slow ones |
| Spilling to disk | When data won’t fit in RAM, the database writes temporary files to storage |
📖 What Apple Actually Shipped
The MacBook Neo dropped March 11, 2026. Here’s what $599 gets you:
- Chip: Apple A18 Pro — 6-core CPU (2 performance, 4 efficiency), 5-core GPU
- RAM: 8 GB unified memory. That’s it. No upgrade option
- Storage: 256 GB base ($599) or 512 GB ($699). Also no upgrade
- Display: 13-inch Liquid Retina, 2408×1506, 500 nits
- Weight: 2.7 lbs. Two USB-C ports. No MagSafe
- Colors: Silver, Indigo, Blush, Citrus
- Battery: 16 hours claimed. EU box doesn’t include a charger
This is the first Mac powered by iPhone silicon. Apple claims 50% faster than Intel Core Ultra 5 laptops and 3x faster on “AI workloads.” Education pricing drops to $499.
📊 The Benchmark That Started the Argument
DuckDB v1.5.0 ran ClickBench (43 queries, 100M rows, ~14 GB Parquet) on three machines:
| Machine | Cores | RAM | Cold Median | Cold Total | Hot Median | Hot Total |
|---|---|---|---|---|---|---|
| MacBook Neo | 6 (2P+4E) | 8 GB | 0.57s | 59.73s | 0.41s | 54.27s |
| AWS c6a.4xlarge | 16 vCPU | 32 GB | 1.34s | 145.08s | 0.50s | 47.86s |
| AWS c8g.metal-48xl | 192 vCPU | 384 GB | 1.54s | 169.67s | 0.05s | 4.35s |
Cold runs: The Neo crushed both cloud instances. Sub-second median. Under a minute total. The AWS boxes with their network-attached storage couldn’t keep up on first reads.
Hot runs: Different story. The 192-core Graviton monster finished in 4.35 seconds — 12.5x faster than the Neo. Even the mid-tier 16-core box edged ahead on total time.
But here’s the thing nobody mentions: the Neo was only 13% slower on hot total runtime than a machine with 10 more CPU threads and 4x the RAM. Dollar for dollar, that’s absurd.
🔍 The Harder Benchmark (TPC-DS)
DuckDB also ran TPC-DS, which is closer to real analytical workloads with joins across 24 tables and 99 queries:
Scale Factor 100 (6 GB memory limit):
- Median query: 1.63 seconds
- Total runtime: 15.5 minutes
Scale Factor 300 (6 GB memory limit):
- Median query: 6.90 seconds
- Query 67 alone: 51 minutes (spilling ~80 GB to disk)
- Total runtime: 79 minutes
At SF300, you’re asking a $599 laptop with 8 GB RAM to chew through a dataset that requires 80 GB of temporary disk writes. It finishes. It just takes a while.
DuckDB’s team set the memory limit to 5 GB (not 8) to avoid fighting macOS for swap space. That’s a common trick — let the database manage its own memory rather than trusting the OS page cache.
🗣️ What People Are Actually Saying
The HN thread split into three camps:
The Believers: One dev replaced 400 lines of C# with 10 lines of DuckDB in an AWS Lambda. Another ported a legacy Python app to Polars/DuckDB and got a 40-80x speedup in two days.
The Skeptics: “Slack and a browser will bring it to its knees” — 8 GB is tight when you’re also running dev tools. Docker containers on 8 GB configs cause noticeable performance hits.
The Realists: “The important question is whether the Neo displaces $350 Windows machines or $1,300 MacBook Airs.” The benchmark story is neat, but the market story matters more.
Multiple commenters also questioned whether 14-75 GB datasets qualify as “big data” in 2026. Fair point. But that’s also the range where most actual analyst work happens.
⚙️ Why Cold Runs Matter More Than You Think
The Neo’s cold-run dominance isn’t a fluke — it reveals a real architectural gap.
Cloud VMs use network-attached EBS volumes. First-read latency on a cold EBS gp3 volume runs 1-3ms per 4KB block. The Neo’s local NVMe? Under 0.1ms. That’s 10-30x faster on random reads.
For analysts who spin up a query, get results, and move on — cold performance IS the performance. You’re not running the same 43 queries in a loop. You’re running one query, reading the answer, then writing another.
The hot-run case (where the 192-core monster dominates) assumes the OS has cached everything in RAM. That requires 384 GB of memory and a stable workload. Most real analytical sessions don’t look like that.
Cool. A $599 laptop runs analytical SQL. Now What the Hell Do We Do? ( ͡° ͜ʖ ͡°)
💰 Sell 'Local-First Analytics' Consulting to SMBs
Small businesses pay $200-800/month for cloud data warehouses they barely use. Most of their datasets are under 50 GB.
Show them DuckDB on a cheap MacBook (or any laptop) can run the same queries faster on cold reads. Set up automated Parquet exports from their SaaS tools. Charge $2-5K for the migration plus $500/month for maintenance.
Example: A freelance data consultant in Portugal migrated 3 small e-commerce clients from BigQuery to DuckDB running on $800 Mac Minis. Saved each client ~$400/month in cloud costs. Charges $1,500/month across all three for ongoing support.
Timeline: 2-3 weeks to build a template pipeline. First client within a month if you’re already doing data work.
🔧 Build a 'Portable Data Lab' Product for Field Researchers
Scientists, field biologists, and NGO workers often collect data in places with unreliable internet. They need to run analysis locally.
Package a pre-configured DuckDB environment with Jupyter notebooks on a cheap MacBook Neo. Bundle domain-specific query templates. Sell it as a turnkey “field analytics station” to research grants and NGOs.
Example: A marine biology PhD student in Indonesia built a portable water quality analysis kit using DuckDB + Python on a refurbished MacBook. Sold the setup guide and template notebooks to 4 other research groups at $800 each. Now consulting for an ocean conservation NGO at $3K/month.
📊 Create a 'Benchmark-as-Content' Newsletter for Data Engineers
DuckDB’s post generated massive engagement because it had real numbers comparing real hardware. Data engineers are starved for honest benchmarks.
Run monthly benchmarks across different hardware/database combos. Publish a free newsletter. Monetize with sponsored deep-dives from database vendors and cloud providers who want their products tested.
Example: A data engineer in Germany started a Substack benchmarking DuckDB vs. ClickHouse vs. Polars on consumer hardware. Hit 4,000 subscribers in 6 months. Now earns $2,800/month from paid tiers and $1,500/month from two vendor sponsorships.
💼 Flip Cheap MacBooks Into 'Data Analyst Workstations'
The $599 MacBook Neo with DuckDB pre-installed is genuinely more capable for analytical work than most $1,500 Windows laptops running Excel. Most people don’t know this.
Buy MacBook Neos at education pricing ($499). Pre-install DuckDB, Jupyter, common data connectors. Sell them as “analyst-ready workstations” to bootcamp graduates and career switchers for $899-999.
Example: A tech reseller in Nigeria buys refurbished MacBooks, pre-installs a data science stack (DuckDB, Python, VS Code, sample datasets), and sells them to data analytics bootcamp graduates for a 40% markup. Moves 8-12 units monthly.
🛠️ Follow-Up Actions
| Step | Action | Resource |
|---|---|---|
| 1 | Install DuckDB and run ClickBench locally | DuckDB docs |
| 2 | Study the actual benchmark scripts | ClickBench DuckDB implementation |
| 3 | Compare your current cloud spend vs. local DuckDB | DuckDB performance guide |
| 4 | Build a sample Parquet pipeline for a real dataset | DuckDB Parquet import guide |
| 5 | Join the DuckDB Discord for community benchmarks | DuckDB community |
Quick Hits
| Want to… | Do this |
|---|---|
 Test DuckDB yourself |
brew install duckdb and run ClickBench locally — takes 20 minutes |
 Cut cloud analytics costs |
Export your warehouse to Parquet, query it with DuckDB on any laptop |
| Benchmark your own hardware |
Fork the ClickBench repo and run the DuckDB implementation |
 Handle datasets > RAM |
Set memory_limit to 60-70% of your RAM and let DuckDB spill to disk |
 Run on even less hardware |
DuckDB ran TPC-H SF100 on an iPhone 16 Pro in under 10 minutes |
A $599 laptop with a phone chip just humiliated a cloud server on first read. The cloud’s not dead — but the bill might be optional.
!