Repository navigation
Expand file tree
/
Copy pathai-supercycle.html
More file actions
1691 lines (1580 loc) · 117 KB
/
Copy pathai-supercycle.html
File metadata and controls
1691 lines (1580 loc) · 117 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>Economics of the AI Supercycle</title>
<!-- KaTeX for math -->
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.16.9/dist/katex.min.css" crossorigin="anonymous">
<script defer src="https://cdn.jsdelivr.net/npm/katex@0.16.9/dist/katex.min.js" crossorigin="anonymous"></script>
<script defer src="https://cdn.jsdelivr.net/npm/katex@0.16.9/dist/contrib/auto-render.min.js" crossorigin="anonymous"
onload="renderMathInElement(document.body, {delimiters:[{left:'$$',right:'$$',display:true}], throwOnError:false});"></script>
<style>
*, *::before, *::after { box-sizing: border-box; margin: 0; padding: 0; }
body {
font-family: -apple-system, BlinkMacSystemFont, "Inter", "Segoe UI", sans-serif;
background: #f5f5f5;
color: #1a1a1a;
line-height: 1.65;
padding: 40px 20px 80px;
}
.page-title {
text-align: center;
font-size: 1.5rem;
font-weight: 700;
margin-bottom: 6px;
color: #111;
}
.page-subtitle {
text-align: center;
font-size: 0.9rem;
color: #777;
margin-bottom: 32px;
}
.container { flex: 1; min-width: 0; display: flex; flex-direction: column; gap: 20px; }
/* ── Sidebar bookmark layout ── */
.layout { display: flex; align-items: flex-start; max-width: 1180px; margin: 0 auto; gap: 28px; }
.sidebar {
flex: 0 0 248px;
position: sticky;
top: 24px;
max-height: calc(100vh - 48px);
overflow-y: auto;
background: #fff;
border-radius: 14px;
box-shadow: 0 1px 4px rgba(0,0,0,.06), 0 4px 16px rgba(0,0,0,.06);
padding: 18px;
}
.sidebar-title { font-size: 0.7rem; text-transform: uppercase; letter-spacing: .1em; color: #888; font-weight: 700; margin-bottom: 14px; }
.sidebar nav ol { list-style: none; padding: 0; margin: 0; }
.sidebar nav > ol > li { margin-bottom: 4px; }
.sidebar nav > ol > li > a {
display: block; font-weight: 700; color: #1a1a1a; text-decoration: none;
padding: 5px 0; line-height: 1.35; font-size: 0.85rem;
}
.sidebar nav > ol > li > a:hover { color: #1d7a45; }
.sidebar nav > ol > li > a.active { color: #1d7a45; }
.sidebar nav ul { list-style: none; padding: 2px 0 6px; margin: 0; }
.sidebar nav ul li a {
display: block; padding: 3px 0 3px 12px; color: #777; text-decoration: none;
font-size: 0.79rem; line-height: 1.4; border-left: 2px solid #eee;
}
.sidebar nav ul li a:hover { color: #1a1a1a; border-left-color: #bbb; }
.sidebar nav ul li a.active { color: #1d7a45; border-left-color: #1d7a45; font-weight: 600; }
html { scroll-behavior: smooth; }
.chapter, .chapter-body h2[id] { scroll-margin-top: 18px; }
@media (max-width: 900px) {
.layout { flex-direction: column; gap: 0; }
.sidebar { position: static; flex: none; width: 100%; max-height: none; margin-bottom: 20px; }
}
/* ── Chapter card ── */
.chapter { background: #fff; border-radius: 14px; box-shadow: 0 1px 4px rgba(0,0,0,.06), 0 4px 16px rgba(0,0,0,.06); overflow: hidden; }
.chapter-header { display: flex; align-items: center; justify-content: space-between; padding: 20px 24px; cursor: pointer; user-select: none; }
.chapter-header-left { display: flex; align-items: center; gap: 16px; }
.chapter-num {
display: flex; align-items: center; justify-content: center;
width: 40px; height: 40px; background: #1a1a1a; color: #fff;
font-weight: 700; font-size: 1rem; border-radius: 8px; flex-shrink: 0;
}
.chapter-title-group { display: flex; flex-direction: column; gap: 2px; }
.chapter-title { font-size: 1.05rem; font-weight: 700; color: #111; }
.chapter-label { font-size: 0.68rem; font-weight: 700; letter-spacing: .1em; text-transform: uppercase; color: #888; }
.chevron { font-size: 1rem; color: #888; transition: transform .25s ease; }
.chapter.collapsed .chevron { transform: rotate(180deg); }
/* ── Chapter body ── */
.chapter-body { padding: 4px 24px 28px; border-top: 1px solid #f0f0f0; }
.chapter.collapsed .chapter-body { display: none; }
.chapter-body p { font-size: 0.97rem; color: #333; margin-bottom: 16px; }
.chapter-body h2 { font-size: 1.05rem; font-weight: 700; margin: 28px 0 12px; color: #111; }
.chapter-body h3 { font-size: 1rem; font-weight: 700; margin: 22px 0 10px; color: #111; }
.chapter-body ul { padding-left: 22px; margin-bottom: 16px; }
.chapter-body ul li { font-size: 0.97rem; color: #333; margin-bottom: 10px; }
.chapter-body ol { padding-left: 26px; margin-bottom: 16px; }
.chapter-body ol li { font-size: 0.97rem; color: #333; margin-bottom: 12px; }
/* ── Inline code badge ── */
code {
font-family: "SFMono-Regular", Consolas, "Liberation Mono", Menlo, monospace;
font-size: 0.82em; background: #f0f0f0; border: 1px solid #ddd; border-radius: 5px;
padding: 1px 7px; color: #333; white-space: nowrap;
}
/* ── Table ── */
.table-wrap { overflow-x: auto; margin: 16px 0 20px; border-radius: 8px; border: 1px solid #e0e0e0; }
table { width: 100%; border-collapse: collapse; font-size: 0.92rem; }
thead tr { background: #1a1a1a; color: #fff; }
thead th { padding: 13px 16px; text-align: left; font-weight: 600; font-size: 0.88rem; }
tbody tr { border-top: 1px solid #e8e8e8; }
tbody tr:nth-child(even) { background: #fafafa; }
tbody td { padding: 12px 16px; color: #333; vertical-align: top; }
/* ── Collapsible sub-section ── */
.sub-collapse { border: 1px solid #e8e8e8; border-radius: 10px; overflow: hidden; margin-top: 20px; }
.sub-collapse-header {
display: flex; align-items: center; justify-content: space-between; padding: 14px 18px;
cursor: pointer; font-size: 0.95rem; font-weight: 600; color: #222; background: #fff; user-select: none;
}
.sub-collapse-header .chevron-orange { color: #e07a20; font-size: 1rem; transition: transform .25s; }
.sub-collapse.open .chevron-orange { transform: rotate(180deg); }
.sub-collapse-body { display: none; padding: 14px 18px 18px; border-top: 1px solid #f0f0f0; font-size: 0.92rem; color: #444; }
.sub-collapse.open .sub-collapse-body { display: block; }
.sub-collapse-body ul { padding-left: 20px; }
.sub-collapse-body ul li { margin-bottom: 8px; }
.sub-collapse-body p { font-size: 0.92rem; margin-bottom: 12px; }
/* ── Chapter footer ── */
.chapter-footer { display: flex; align-items: center; justify-content: flex-end; gap: 16px; padding: 16px 24px 24px; }
.btn-collapse { background: none; border: none; cursor: pointer; font-size: 0.88rem; color: #555; display: flex; align-items: center; gap: 5px; font-weight: 500; }
.btn-next {
background: #1a1a1a; color: #fff; border: none; border-radius: 999px; padding: 10px 22px;
font-size: 0.9rem; font-weight: 600; cursor: pointer; display: flex; align-items: center; gap: 6px; transition: background .15s;
}
.btn-next:hover { background: #333; }
/* ── TIP callout ── */
.tip-box { background: #f0faf4; border-left: 3px solid #34c76f; border-radius: 8px; padding: 16px 18px; margin: 18px 0; }
.tip-label { font-size: 0.75rem; font-weight: 700; letter-spacing: .08em; color: #22a05a; margin-bottom: 6px; display: flex; align-items: center; gap: 6px; font-style: italic; }
.tip-box p { font-size: 0.92rem; font-style: italic; color: #2d4a38; margin: 0 !important; }
.tip-box p + p { margin-top: 10px !important; }
/* ── WARNING / DEBATE callout ── */
.warn-box { background: #fff8f0; border-left: 3px solid #e8a33d; border-radius: 8px; padding: 16px 18px; margin: 18px 0; }
.warn-label { font-size: 0.75rem; font-weight: 700; letter-spacing: .08em; color: #c87f1e; margin-bottom: 6px; display: flex; align-items: center; gap: 6px; font-style: italic; }
.warn-box p { font-size: 0.92rem; color: #5c4626; margin: 0 !important; }
.warn-box p + p { margin-top: 10px !important; }
/* ── Quote box ── */
.quote-box { border-left: 3px solid #c9c9c9; background: #fafafa; border-radius: 0 8px 8px 0; padding: 14px 18px; margin: 18px 0; }
.quote-box p { font-size: 0.97rem; font-style: italic; color: #2a2a2a; margin: 0 0 6px !important; }
.quote-attr { font-size: 0.8rem; color: #888; font-weight: 600; font-style: normal; }
/* ── Stat grid ── */
.stat-grid { display: grid; grid-template-columns: repeat(auto-fit, minmax(150px, 1fr)); gap: 12px; margin: 18px 0 20px; }
.stat-card { background: #fafafa; border: 1px solid #e8e8e8; border-radius: 10px; padding: 14px 16px; }
.stat-num { font-size: 1.5rem; font-weight: 800; color: #1d7a45; line-height: 1.1; }
.stat-label { font-size: 0.74rem; color: #777; margin-top: 6px; line-height: 1.35; }
/* ── Layer-cake / triangle visual ── */
.cake { margin: 20px 0; display: flex; flex-direction: column; gap: 6px; }
.cake-row {
display: flex; align-items: center; gap: 12px; border-radius: 8px; padding: 12px 16px;
color: #fff; font-size: 0.9rem;
}
.cake-row .cake-name { font-weight: 700; min-width: 92px; }
.cake-row .cake-desc { font-size: 0.84rem; opacity: 0.92; }
.cake-apps { background: #2b8a5b; }
.cake-model { background: #3a7bd5; }
.cake-infra { background: #6c5ce7; }
.cake-power { background: #d98324; }
.cake-semis { background: #1a1a1a; }
.cake-caption { text-align: center; font-size: 0.86rem; color: #555; font-style: italic; margin-top: 12px; }
/* ── Cost bar (electrons-to-tokens) ── */
.costbar-wrap { margin: 16px 0 8px; }
.costbar { display: flex; align-items: center; gap: 10px; margin: 7px 0; font-size: 0.86rem; }
.costbar-label { min-width: 168px; color: #333; }
.costbar-track { flex: 1; background: #f0f0f0; border-radius: 5px; overflow: hidden; height: 22px; }
.costbar-fill { height: 100%; background: #3a7bd5; border-radius: 5px; }
.costbar-val { min-width: 96px; text-align: right; font-weight: 600; color: #222; font-variant-numeric: tabular-nums; }
/* ── Horizontal bar chart ── */
.chart-title { font-size: 0.82rem; font-weight: 700; color: #555; text-transform: uppercase; letter-spacing: .05em; margin: 20px 0 2px; }
.chart-src { font-size: 0.74rem; color: #aaa; font-style: italic; margin: 2px 0 12px; }
.hbar-chart { margin: 12px 0 8px; }
.hbar-row { display: flex; align-items: center; gap: 10px; margin: 7px 0; font-size: 0.85rem; }
.hbar-label { min-width: 150px; color: #333; }
.hbar-track { flex: 1; background: #f0f0f0; border-radius: 5px; overflow: hidden; height: 24px; }
.hbar-fill { height: 100%; border-radius: 5px; min-width: 2px; }
.hbar-val { min-width: 104px; text-align: right; font-weight: 600; color: #222; font-variant-numeric: tabular-nums; }
.hbar-semis { background: #1a1a1a; } .hbar-infra { background: #6c5ce7; } .hbar-apps { background: #2b8a5b; }
.hbar-blue { background: #3a7bd5; } .hbar-amber { background: #d98324; } .hbar-grey { background: #9aa0a6; }
.hbar-row.dim .hbar-fill { opacity: .55; }
/* ── Stacked share bar ── */
.share-bar { display: flex; height: 36px; border-radius: 8px; overflow: hidden; margin: 14px 0 8px; font-size: 0.78rem; font-weight: 700; color: #fff; }
.share-seg { display: flex; align-items: center; justify-content: center; min-width: 0; }
.share-legend { display: flex; flex-wrap: wrap; gap: 16px; font-size: 0.8rem; color: #555; margin-bottom: 14px; }
.share-legend span { display: flex; align-items: center; gap: 6px; }
.share-dot { width: 11px; height: 11px; border-radius: 3px; display: inline-block; flex-shrink: 0; }
/* ── Value-accrual funnel (the inverted-triangle chart) ── */
.accrue { display: flex; gap: 28px; flex-wrap: wrap; justify-content: center; margin: 16px 0 6px; }
.accrue-col { flex: 1; min-width: 280px; }
.accrue-head { text-align: center; font-weight: 700; font-size: 0.92rem; color: #111; margin-bottom: 14px; }
.accrue-head span { font-weight: 500; color: #888; font-size: 0.8rem; }
.funnel { display: flex; flex-direction: column; gap: 7px; }
.funnel-row { display: flex; align-items: center; gap: 10px; }
.funnel-layer { width: 48px; font-size: 0.76rem; color: #666; text-align: right; flex-shrink: 0; }
.funnel-track { flex: 1; display: flex; justify-content: center; }
.funnel-bar {
height: 44px; display: flex; align-items: center; justify-content: center;
color: #0b3a56; font-weight: 700; font-size: 0.84rem; border-radius: 6px;
background: #8ec9ef; min-width: 44px; box-shadow: inset 0 0 0 1px rgba(0,0,0,.06);
}
.funnel-cloud .funnel-bar { background: #b9d9c9; color: #14402c; }
.accrue-note { text-align: center; font-size: 0.8rem; color: #555; margin-top: 12px; }
.accrue-note strong { color: #1d7a45; }
.funnel-cloud .accrue-note strong { color: #14402c; }
/* ── Source figure (original chart image) ── */
.fig { margin: 18px 0 10px; }
.fig img { width: 100%; height: auto; display: block; border: 1px solid #e4e4e4; border-radius: 10px; background: #fff; }
.fig-cap { text-align: center; font-size: 0.84rem; color: #666; font-style: italic; margin-top: 9px; }
.fig-2 { display: grid; grid-template-columns: 1fr 1fr; gap: 16px; align-items: start; }
@media (max-width: 760px) { .fig-2 { grid-template-columns: 1fr; } }
/* ── Two-column compare (IQ vs EQ) ── */
.vs2 { display: grid; grid-template-columns: 1fr 1fr; gap: 14px; margin: 16px 0 10px; }
@media (max-width: 700px) { .vs2 { grid-template-columns: 1fr; } }
.vs-card { border: 1px solid #e8e8e8; border-radius: 10px; padding: 14px 16px; }
.vs-card h4 { font-size: 0.9rem; font-weight: 700; margin: 0 0 4px; display: flex; align-items: center; gap: 8px; }
.vs-card .vs-sub { font-size: 0.78rem; color: #888; margin-bottom: 10px; }
.vs-card ul { padding-left: 18px; margin: 0; }
.vs-card li { font-size: 0.9rem; margin-bottom: 7px; color: #333; }
.vs-down { background: #fff6f5; } .vs-down h4 { color: #b5462f; }
.vs-up { background: #f3fbf6; } .vs-up h4 { color: #1d7a45; }
.vs-tag { font-size: 0.64rem; font-weight: 700; letter-spacing: .06em; text-transform: uppercase; border-radius: 999px; padding: 2px 9px; }
.vs-down .vs-tag { background: #f6d9d2; color: #b5462f; } .vs-up .vs-tag { background: #cdeed9; color: #1d7a45; }
.speaker-tag {
display: inline-block; font-size: 0.72rem; font-weight: 700; letter-spacing: .04em;
background: #eef3ff; color: #3a5bd5; border: 1px solid #d6e0ff; border-radius: 999px;
padding: 2px 10px; margin-bottom: 14px;
}
.src-link { font-size: 0.8rem; color: #999; }
.src-link a { color: #1d7a45; text-decoration: none; }
.src-link a:hover { text-decoration: underline; }
.katex { font-size: 1em; }
/* ── Save-as-PDF button ── */
.pdf-actions { text-align: center; margin: -18px 0 30px; }
.btn-pdf {
display: inline-flex; align-items: center; gap: 8px;
background: #1a1a1a; color: #fff; border: none; border-radius: 999px;
padding: 9px 20px; font-size: 0.85rem; font-weight: 600; cursor: pointer;
font-family: inherit; transition: background .15s, opacity .15s;
box-shadow: 0 1px 4px rgba(0,0,0,.12);
}
.btn-pdf:hover { background: #333; }
.btn-pdf[disabled] { opacity: .6; cursor: progress; }
/* ── Print / PDF stylesheet ── */
@media print {
@page { margin: 14mm 12mm; }
* { -webkit-print-color-adjust: exact !important; print-color-adjust: exact !important; }
html { scroll-behavior: auto; }
body { background: #fff; padding: 0; font-size: 11pt; line-height: 1.5; }
.page-title { font-size: 20pt; margin-bottom: 4px; }
.page-subtitle { margin-bottom: 18px; }
.sidebar, .pdf-actions, .chapter-footer, .chevron,
.btn-collapse, .btn-next, .sub-collapse-header .chevron-orange { display: none !important; }
.layout { display: block; max-width: none; margin: 0; gap: 0; }
.container { gap: 0; }
/* show everything — nothing hidden behind collapse */
.chapter.collapsed .chapter-body { display: block !important; }
.sub-collapse-body { display: block !important; }
/* each chapter starts on a fresh page (except the first) */
.chapter { break-before: page; box-shadow: none; border: 1px solid #ccc; border-radius: 8px; margin-bottom: 12px; }
#ch1 { break-before: avoid; }
.chapter-header { break-inside: avoid; break-after: avoid; cursor: default; }
/* keep visual units from splitting across pages */
.fig, .fig-2, .table-wrap, table, img, .tip-box, .warn-box, .quote-box,
.stat-grid, .stat-card, .vs2, .vs-card, .cake, .accrue, .hbar-chart,
.share-bar, .sub-collapse, .costbar-wrap { break-inside: avoid; }
h2, h3, .chart-title { break-after: avoid; }
p, li { orphans: 3; widows: 3; }
a { color: #1d7a45; text-decoration: none; }
.fig img { max-width: 100%; height: auto; }
}
/* ── Era timeline (mobile vs AI) ── */
.tl { border: 1px solid #e6e6e6; border-radius: 10px; overflow: hidden; margin: 16px 0 10px; }
.tl-row { display: grid; grid-template-columns: 168px 1fr 1fr; }
.tl-row + .tl-row { border-top: 1px solid #eee; }
.tl-h { background: #1a1a1a; color: #fff; font-weight: 700; font-size: 0.8rem; padding: 11px 14px; }
.tl-stage { font-weight: 700; font-size: 0.84rem; padding: 13px 14px; background: #fafafa; color: #111; display: flex; align-items: center; }
.tl-cell { padding: 13px 14px; font-size: 0.85rem; border-left: 1px solid #eee; color: #333; }
.tl-year { display: block; font-weight: 700; font-size: 0.78rem; color: #3a7bd5; margin-bottom: 2px; }
.tl-now { background: #f3fbf6; }
.tl-now .tl-year { color: #1d7a45; }
.tl-here { display: inline-block; margin-top: 6px; font-size: 0.7rem; font-weight: 700; color: #1d7a45; background: #cdeed9; border-radius: 999px; padding: 2px 9px; }
@media (max-width: 700px) { .tl-row { grid-template-columns: 1fr; } .tl-cell { border-left: none; border-top: 1px solid #f0f0f0; } }
</style>
</head>
<body>
<div class="page-title">Economics of the AI Supercycle</div>
<div class="page-subtitle">Reading notes · Stanford MS&E 435, Spring ’26 (Apoorv Agrawal & guests) · Classes 1–4 + companion essays</div>
<div class="pdf-actions">
<button class="btn-pdf" id="pdfBtn" onclick="exportPDF()">📄 Save as PDF</button>
</div>
<div class="layout">
<aside class="sidebar">
<div class="sidebar-title">Contents</div>
<nav>
<ol>
<li>
<a href="#ch1">1. Where’s the Money in AI?</a>
<ul>
<li><a href="#ch1-triangle">The inverted triangle</a></li>
<li><a href="#ch1-cake">The five-layer cake</a></li>
<li><a href="#ch1-marginal">Why AI ≠ cloud</a></li>
<li><a href="#ch1-cycles">Lessons from past supercycles</a></li>
<li><a href="#ch1-profit">Profit by layer</a></li>
<li><a href="#ch1-scorecard">The 2026 scorecard</a></li>
<li><a href="#ch1-silicon">The custom-silicon wildcard</a></li>
<li><a href="#ch1-consumer">Consumer AI & monetization</a></li>
<li><a href="#ch1-traininf">Training vs inference</a></li>
</ul>
</li>
<li>
<a href="#ch2">2. The Economics of Inference</a>
<ul>
<li><a href="#ch2-marginal">The end of zero marginal cost</a></li>
<li><a href="#ch2-token">The token as atomic unit</a></li>
<li><a href="#ch2-groq">Groq & deterministic compute</a></li>
<li><a href="#ch2-deal">The Nvidia–Groq deal</a></li>
<li><a href="#ch2-cost">Why inference keeps getting cheaper</a></li>
<li><a href="#ch2-bubble">The bubble debate</a></li>
<li><a href="#ch2-agents">Agents, harnesses & safety</a></li>
</ul>
</li>
<li>
<a href="#ch3">3. From Electrons to Tokens</a>
<ul>
<li><a href="#ch3-capex">The CapEx supercycle</a></li>
<li><a href="#ch3-diglabor">Digital labor</a></li>
<li><a href="#ch3-energy">The energy-first thesis</a></li>
<li><a href="#ch3-cost">What $60M/MW buys</a></li>
<li><a href="#ch3-revenue">Making money: payback math</a></li>
<li><a href="#ch3-longshort">Long / short the stack</a></li>
</ul>
</li>
<li>
<a href="#ch4">4. We Already Have AGI</a>
<ul>
<li><a href="#ch4-agi">The provocation</a></li>
<li><a href="#ch4-context">The context problem</a></li>
<li><a href="#ch4-dynamo">Dynamo to computer</a></li>
<li><a href="#ch4-connector">The connector story</a></li>
<li><a href="#ch4-software">Is software dead?</a></li>
<li><a href="#ch4-value">Where value accrues</a></li>
<li><a href="#ch4-advice">Career advice</a></li>
</ul>
</li>
<li>
<a href="#ch5">5. The State of Consumer AI</a>
<ul>
<li><a href="#ch5-usage">Usage: the 20× run</a></li>
<li><a href="#ch5-tiers">App tiers & stock vs flow</a></li>
<li><a href="#ch5-seasonal">Threats that faded</a></li>
<li><a href="#ch5-engagement">Engagement & retention</a></li>
<li><a href="#ch5-smile">The smile curve</a></li>
<li><a href="#ch5-time">Time is money</a></li>
<li><a href="#ch5-ads">Ads vs subscription</a></li>
</ul>
</li>
<li>
<a href="#ch6">6. Where Does Value Go From Here?</a>
<ul>
<li><a href="#ch6-stack">Value migrates up the stack</a></li>
<li><a href="#ch6-timeline">Mobile vs AI: what inning?</a></li>
<li><a href="#ch6-native">The native company isn’t born</a></li>
<li><a href="#ch6-counter">The counter-argument</a></li>
<li><a href="#ch6-bottom">Bottom line</a></li>
</ul>
</li>
</ol>
</nav>
</aside>
<div class="container">
<!-- ════════════════ CHAPTER 1 ════════════════ -->
<div class="chapter" id="ch1">
<div class="chapter-header" onclick="toggleChapter('ch1')">
<div class="chapter-header-left">
<div class="chapter-num">1</div>
<div class="chapter-title-group">
<div class="chapter-title">💰 Where’s the Money in AI?</div>
<div class="chapter-label">Class 1 · Apoorv Agrawal (Altimeter)</div>
</div>
</div>
<span class="chevron">⌃</span>
</div>
<div class="chapter-body">
<span class="speaker-tag">Lecturer: Apoorv Agrawal · Partner, Altimeter Capital</span>
<p>
The opening lecture frames the entire course around one question: <strong>the AI build-out is
consuming hundreds of billions in capital — where does the economic value actually
land?</strong> Apoorv leads the AI practice at Altimeter, a concentrated public + private
investment firm that “repositioned its entire focus” around this supercycle. His
claim is that this is the biggest technology cycle yet — bigger than internet, mobile,
or cloud — and that in five years everyone will ask you, <em>“Did you see it
coming?”</em>
</p>
<h2 id="ch1-triangle">The inverted triangle</h2>
<p>
The signature slide is a value-accrual map of the AI ecosystem. In cloud and software, the
money pooled at the <em>top</em> — the application layer — producing a healthy
pyramid. In AI today the shape is <strong>inverted</strong>: value concentrates at the
<em>bottom</em> of the stack (silicon), and the application layer is comparatively tiny.
</p>
<div class="fig">
<img loading="lazy" alt="Where does value accrue in the Gen AI stack? Cloud (apps $400B at top) vs Generative AI inverted triangle (semis $75B at bottom), Q1 2024" src="https://substackcdn.com/image/fetch/$s_!CADs!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F48e9c94a-881a-4148-945f-8475bb485737_2048x955.png">
<div class="fig-cap">The signature contrast (Q1 2024): in cloud the money pools at <em>apps</em> (top); in generative AI it pools at <em>semis</em> (bottom) — an inverted triangle. Source: Apoorv Agrawal / Altimeter (@apoorv03).</div>
</div>
<div class="fig">
<img loading="lazy" alt="Where does value accrue in AI? Q1 2026 estimated annual revenue by layer" src="https://substackcdn.com/image/fetch/$s_!T85K!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0310bf44-2cfe-403a-a702-62cc7e499f48_2030x926.png">
<div class="fig-cap">Two years and ~5× growth later (Q1 2026), the shape barely changed: apps ~$60B, infra ~$75B, semis ~$300B. Source: Apoorv Agrawal / Altimeter.</div>
</div>
<div class="tip-box">
<div class="tip-label">💡 THE CORE PUZZLE</div>
<p>In two years the AI ecosystem grew <strong>~5×, from ~$90B to ~$435B</strong> in annualized revenue — yet the <em>shape</em> barely changed. Semis still take <strong>~79% of gross profit</strong>; the app layer grew 12× but is still tiny. The central bet of the course: will it flip toward the cloud/software shape — in 5 years, 10, 15, or never? (Apoorv now estimates <em>well over a decade</em>.)</p>
</div>
<p>Students proposed several theories for why AI looks so different from cloud, and Apoorv endorsed all of them as partial truths:</p>
<ul>
<li><strong>It’s early.</strong> The cycle is young; triangles take years to flip.</li>
<li><strong>Nvidia has a stranglehold.</strong> One player “runs the tables” on the most profitable layer.</li>
<li><strong>The physics are different.</strong> Inference is genuinely expensive to run, unlike near-free software distribution.</li>
</ul>
<h2 id="ch1-cake">The five-layer cake</h2>
<p>
Borrowing Jensen Huang’s framing, the AI stack is a “five-layer cake.” A data
center turns power and chips into capacity you can rent <em>by the hour</em> (to train) or
<em>by the token</em> (to serve).
</p>
<div class="cake">
<div class="cake-row cake-apps"><span class="cake-name">Apps</span><span class="cake-desc">ChatGPT, Gemini, Claude, coding agents, vertical SKUs (Salesforce Einstein, Palantir AIP)</span></div>
<div class="cake-row cake-model"><span class="cake-name">Models</span><span class="cake-desc">OpenAI, Anthropic, Gemini, xAI — the labs training frontier weights</span></div>
<div class="cake-row cake-infra"><span class="cake-name">Infra</span><span class="cake-desc">Inference / cloud layer — most competitive, highest “metabolic rate,” least stable equilibrium</span></div>
<div class="cake-row cake-power"><span class="cake-name">Power</span><span class="cake-desc">Energy, interconnects, memory — the physical substrate</span></div>
<div class="cake-row cake-semis"><span class="cake-name">Semis</span><span class="cake-desc">Chips — Nvidia plus the ASIC programs (TPU, Trainium, MTIA …)</span></div>
</div>
<div class="cake-caption">The stack, top to bottom. Today value pools at the bottom; the bet is that it migrates upward over time.</div>
<p>How Apoorv attributes conglomerates and “old-economy” reinventors:</p>
<ul>
<li><strong>Google</strong> is split by business unit: TPU → semis, GCP → infra, Gemini → apps.</li>
<li><strong>Salesforce / Palantir</strong> AI SKUs are captured in the app layer “by way of the substrate” — their model/inference spend shows up there even if it’s hard to extract from disclosures.</li>
</ul>
<h2 id="ch1-marginal">Why AI isn’t cloud</h2>
<p>
Marc Andreessen’s “software ate the world” worked because the marginal cost of
another user was ~zero — SaaS ran at 80–90% gross margins. AI breaks that model:
<strong>the incremental AI user is not free</strong> because you have to “burn the
GPUs.” This is why large consumer-AI businesses can still be unprofitable at billions in
revenue.
</p>
<h2 id="ch1-cycles">Lessons from past supercycles</h2>
<p>
Triangles flip slowly. AWS is the canonical example of how long the CapEx-to-value lag runs:
</p>
<div class="table-wrap">
<table>
<thead><tr><th>Milestone</th><th>Year</th><th>Note</th></tr></thead>
<tbody>
<tr><td>AWS founded / first CapEx</td><td>2004</td><td>“Is Amazon going to go bankrupt?” was the loudest question on earnings calls</td></tr>
<tr><td>First external customer (Netflix)</td><td>2010</td><td>~6 years to a marquee tenant</td></tr>
<tr><td>Amazon fully shifts to AWS</td><td>2012</td><td><strong>~8 years</strong> from breaking ground to the model paying off</td></tr>
</tbody>
</table>
</div>
<p>
Apoorv suspects the AI triangle may stay inverted <em>longer</em> than cloud’s ~decade,
because getting the “substrate” right is so hard. Two things could re-price the
bottom layers: (1) a <strong>breakout ASIC program</strong> at a hyperscaler (TPU, MTIA,
Trainium, etc.), or (2) hyperscalers <strong>stopping their big CapEx guidance</strong> —
which would imply the current equilibrium doesn’t work. He recommends listening to
hyperscaler earnings calls four times a year as free CEO-level signal.
</p>
<div class="tip-box">
<div class="tip-label">💡 SUPERCYCLE WINNERS & INTEGRATION</div>
<p>Each prior cycle minted a dominant, vertically-integrated winner — <strong>internet → Google</strong> (~$3T, ~99% search share), <strong>mobile → Apple</strong>, <strong>social → Meta</strong> (“maybe lost a trillion by not going down to the servers”). Cloud was the exception: a heterogeneous oligopoly (AWS / GCP / Azure), nobody fully integrated. The open question for AI: one integrated winner, or another oligopoly?</p>
</div>
<h2 id="ch1-profit">Profitability by layer</h2>
<p>
Viewed through margins, the triangle is <em>even more</em> concentrated than it looks by
revenue. The further <em>up</em> the stack, the thinner the margin — so even as revenue
spreads, profit stays pinned to the bottom.
</p>
<div class="stat-grid">
<div class="stat-card"><div class="stat-num">~73%</div><div class="stat-label">Semis gross margin</div></div>
<div class="stat-card"><div class="stat-num">~55%</div><div class="stat-label">Infrastructure gross margin</div></div>
<div class="stat-card"><div class="stat-num">~33%</div><div class="stat-label">Application gross margin</div></div>
<div class="stat-card"><div class="stat-num">~40%</div><div class="stat-label">Share of Nvidia’s fleet used for inference (vs ~60% training)</div></div>
</div>
<div class="fig-2">
<div class="fig">
<img loading="lazy" alt="Gross margins by layer: semis ~73%, infra ~55%, apps ~33%" src="https://substackcdn.com/image/fetch/$s_!F06c!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8df10c62-91e8-457f-bff0-ffd98dc2fe2d_2012x860.png">
<div class="fig-cap">Gross margin compresses up the stack: semis ~73%, infra ~55%, apps ~33%.</div>
</div>
<div class="fig">
<img loading="lazy" alt="Gross-profit share by layer, semis declining from 87% to 79%" src="https://substackcdn.com/image/fetch/$s_!YvZD!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F54b528e3-82a4-4e13-bc21-34822c2b7f37_1664x1336.png">
<div class="fig-cap">Gross-profit share: semis still take ~79% (down only ~8 pts from 87% in 2024).</div>
</div>
</div>
<div class="fig">
<img loading="lazy" alt="Revenue growth by layer over two years, semis adding far more than apps" src="https://substackcdn.com/image/fetch/$s_!v75D!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe5c4d9a9-d4d0-401b-9177-07529e5b6af7_1670x1298.png">
<div class="fig-cap">Incremental revenue added by layer over two years — semis (~$225B) dwarf apps (~$55B). Source: Apoorv Agrawal / Altimeter.</div>
</div>
<p>
The dollar gap is the punchline: over two years the semis layer added <strong>~$225B</strong>
of revenue versus <strong>~$55B</strong> for apps — and <strong>Nvidia alone added ~$175B</strong>,
roughly <strong>3× the entire app layer</strong>. Inside each layer the concentration is
extreme, which Apoorv compresses into one line:
</p>
<div class="tip-box">
<div class="tip-label">💡 ONE LINE TO REMEMBER</div>
<p>“<strong>Semiconductors are a one-player game. Apps are a two-player game. Infrastructure is the only competitive layer.</strong>” Nvidia is ~80% of the ~$300B semis layer (~$250B; Broadcom ~$34B, custom-HBM ~$25B); OpenAI + Anthropic are ~75% of the ~$60B app layer (~$45B); only infra is spread across Azure / AWS / GCP / Oracle ($10–20B each) plus CoreWeave (~$6B).</p>
</div>
<h2 id="ch1-scorecard">The 2026 scorecard: capex and the ROI question</h2>
<p>
The build-out that feeds the bottom of the triangle keeps accelerating. Hyperscaler capex is
the clearest tell:
</p>
<div class="fig">
<img loading="lazy" alt="Hyperscaler capex 2024-2026 rising from ~$256B to ~$443B to ~$600B+" src="https://substackcdn.com/image/fetch/$s_!iBUK!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa33a36fb-d608-478f-9932-d84986186281_1183x559.png">
<div class="fig-cap">Hyperscaler capex: ~$256B (2024) → ~$443B (2025, +73%) → ~$600B+ (2026 est.). Source: Apoorv Agrawal / Altimeter.</div>
</div>
<p>
About <strong>75% (~$450B)</strong> of the 2026 number is aimed at AI infrastructure — and
the unanswered question is the return on it. The hyperscaler CEOs (the “listen four times
a year” signal) are betting yes, with varying candor:
</p>
<ul>
<li><strong>Andy Jassy (Amazon):</strong> “As fast as we’re adding capacity… we’re monetizing it.”</li>
<li><strong>Sundar Pichai (Google):</strong> conceded there are “elements of irrationality” in the spend, while asserting positive ROI.</li>
<li><strong>Mark Zuckerberg (Meta):</strong> argues for aggressively front-loading capacity — the worst case is merely “building new infrastructure slowly” into demand you already have.</li>
</ul>
<h2 id="ch1-silicon">The custom-silicon wildcard</h2>
<p>
Apoorv’s “biggest variable” for whether margins ever compress and profit shifts
upstream is the hyperscalers’ in-house accelerator (ASIC) programs. The competitive
pressure is already real: Anthropic’s order of up to <strong>1 million Google TPUs</strong>
reportedly pushed Nvidia to cut pricing <strong>~30%</strong> for some customers.
</p>
<div class="table-wrap">
<table>
<thead><tr><th>Program</th><th>Status (2026)</th><th>Signal</th></tr></thead>
<tbody>
<tr><td><strong>Google TPU</strong></td><td>7th-gen “Ironwood” GA; sold as merchant hardware</td><td>Most mature; Anthropic ordered up to 1M chips; forced Nvidia price cuts</td></tr>
<tr><td><strong>Amazon Trainium</strong></td><td>1.4M Trainium2 deployed; Trainium3 (3nm) in production</td><td>Custom-chip business ~$10B run-rate, triple-digit growth; powers most Bedrock inference</td></tr>
<tr><td><strong>OpenAI ASIC + AMD</strong></td><td>10 GW Broadcom custom accelerators (from 2026) + 6 GW AMD MI450</td><td>Diversifying off Nvidia at gigawatt scale</td></tr>
<tr><td><strong>Microsoft Maia</strong></td><td>Maia 200 in Azure; claims 3× Trainium3 on inference</td><td>Already powering a portion of ChatGPT workloads</td></tr>
<tr><td><strong>Meta MTIA</strong></td><td>MTIA v3 for internal inference; acquired Rivos (RISC-V)</td><td>Internal-only, but reduces Nvidia dependence</td></tr>
</tbody>
</table>
</div>
<div class="fig">
<img loading="lazy" alt="Cumulative AI compute capacity by chip type in H100-equivalents, Nvidia dominant with custom silicon at the margins" src="https://substackcdn.com/image/fetch/$s_!l_3_!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2b4e875a-1015-4982-b064-24a88a6f7d14_2048x1515.png">
<div class="fig-cap">Cumulative AI compute by chip type (H100-equivalents): Nvidia still overwhelming, with custom silicon emerging at the margins. Source: Apoorv Agrawal / Altimeter.</div>
</div>
<div class="tip-box">
<div class="tip-label">💡 WHY THIS IS THE HINGE</div>
<p>If even one ASIC program has “breakout success” (Apoorv’s Class 1 point), it becomes the single biggest re-pricing event for the semis layer — the most likely way the inverted triangle finally starts to flip. So far, custom silicon shows up only “at the margins,” and Nvidia still adds 3× the app layer in a year.</p>
</div>
<h2 id="ch1-consumer">Consumer AI and the monetization gap</h2>
<p>
Outside coding, consumer AI is the biggest market — but it barely monetizes. Apoorv
benchmarks the leading AI app against the great consumer franchises:
</p>
<div class="table-wrap">
<table>
<thead><tr><th>Franchise</th><th>Users</th><th>ARPU / year</th></tr></thead>
<tbody>
<tr><td>Alphabet</td><td>~4B</td><td>~$100</td></tr>
<tr><td>Meta</td><td>~3.5B</td><td>~$70</td></tr>
<tr><td>ChatGPT (leading AI app)</td><td>~1B (95% free)</td><td><strong>~$10</strong></td></tr>
</tbody>
</table>
</div>
<p>So there are two gaps to close, and Apoorv is skeptical of the obvious answers:</p>
<ul>
<li><strong>1B → 4B users.</strong> Knowledge work alone won’t get there — not everyone online does “active” query work. You’d have to go beyond knowledge work.</li>
<li><strong>$10 → $100 ARPU.</strong> Probably not subscription — likely <strong>ads</strong>. AI ads could price well because of logged-in identity, understood intent, strong attribution, and trust. (Echoes the “ads won’t work on phones” skepticism before mobile ads exploded.)</li>
</ul>
<p>
He also maps AI apps onto a taxonomy of consumer products — <strong>mandatory</strong>
(WhatsApp, Chrome, YouTube; ~3B), <strong>social</strong> (Instagram, TikTok, Facebook;
network effects; ~1.5–2B), and <strong>niche</strong> (Spotify, Amazon, Twitter). Today
ChatGPT has just overtaken the <em>niche</em> tier and is heading toward <em>social</em>;
Gemini trails (with the question of how much is product quality vs Google’s distribution
advantage). As an investor he’d have loved to see it head straight for “mandatory
utility,” but knowledge work may cap it short of that.
</p>
<h2 id="ch1-traininf">Training vs inference</h2>
<p>
Whether the triangle flips may hinge on inference eventually dwarfing training. The
shapes of the two workloads differ sharply:
</p>
<ul>
<li><strong>Training:</strong> predictable, high-utilization, bursts of intense work over a short window.</li>
<li><strong>Inference:</strong> bursty, human-driven (“dips around Thanksgiving and Christmas”), harder to predict — until agents run 24/7.</li>
</ul>
<p>
Nvidia has quoted inference at ~40% of its fleet; Apoorv expects that to drift toward
inference over time, but won’t call when. There’s also a timing mismatch baked into
the cycle: semis are built for 5–6-year horizons while app revenue is “for right
now,” which makes the lower half of the stack cyclical — “like laying down
the railroads.”
</p>
<div class="sub-collapse" id="sub-ch1-quiz">
<div class="sub-collapse-header" onclick="toggleSubCollapse('sub-ch1-quiz')">
<span>🏆 Day-one quiz — and the hidden hint</span>
<span class="chevron-orange">⌃</span>
</div>
<div class="sub-collapse-body">
<p>The class opened with a live quiz naming companies across the stack — answers included <strong>Nvidia</strong>, <strong>CoreWeave</strong>, and <strong>Crusoe</strong>. The hidden hint: every right answer is a CEO scheduled to guest-teach later in the course (Crusoe’s Chase Lochmiller appears in Class 3). Apoorv’s framing for those guests: <em>are you dominant, how long do you stay dominant, what are your pricing-compression vectors, and — for infra startups — are you a feature or a platform? (i.e., “why is this not already part of AWS?”)</em></p>
</div>
</div>
</div>
<div class="chapter-footer">
<button class="btn-collapse" onclick="toggleChapter('ch1')">⌃ Collapse</button>
<button class="btn-next" onclick="scrollToChapter('ch2')">Next Chapter →</button>
</div>
</div>
<!-- ════════════════ CHAPTER 2 ════════════════ -->
<div class="chapter" id="ch2">
<div class="chapter-header" onclick="toggleChapter('ch2')">
<div class="chapter-header-left">
<div class="chapter-num">2</div>
<div class="chapter-title-group">
<div class="chapter-title">⚡ The Economics of Inference</div>
<div class="chapter-label">Class 2 · Brad Gerstner (Altimeter) & Sunny Madra (Nvidia / ex-Groq)</div>
</div>
</div>
<span class="chevron">⌃</span>
</div>
<div class="chapter-body">
<span class="speaker-tag">Guests: Brad Gerstner, Founder/CEO Altimeter · Sunny Madra, ex-President of Groq, now at Nvidia</span>
<p>
Class 2 zooms into the most contested layer: <strong>inference</strong> — the actual
production of tokens. Brad founded Altimeter in 2008 (now ~$15B AUM) and is an investor in
OpenAI, Anthropic, Nvidia, Cerebras, and Groq. Sunny is a serial founder whose company Groq
was acquired by Nvidia for ~$20B — Nvidia’s largest acquisition ever. Together
they unpack why the unit economics of intelligence are flipping from scary to attractive.
</p>
<h2 id="ch2-marginal">The end of zero marginal cost</h2>
<p>
The premise restated: software produced near-zero incremental cost of distribution; AI does
not. More users means more compute. To set scale, Sunny opens with global GDP per capita over
2,000 years — flat for ~1,800 years, then vertical.
</p>
<div class="stat-grid">
<div class="stat-card"><div class="stat-num">~25 yrs</div><div class="stat-label">Time to double global GDP today (was centuries)</div></div>
<div class="stat-card"><div class="stat-num">5% → 13%</div><div class="stat-label">Tech as a share of global GDP — and rising</div></div>
<div class="stat-card"><div class="stat-num">15% vs 6%</div><div class="stat-label">NASDAQ vs non-tech EPS growth (10-yr CAGR)</div></div>
<div class="stat-card"><div class="stat-num">10× / 10×</div><div class="stat-label">Demis Hassabis: 10× the impact of the Industrial Revolution, at 10× the speed</div></div>
</div>
<p>
Brad’s through-line: innovation is itself a societal good (correlated with lower poverty,
higher literacy, more democracy, lower child mortality), and AI accelerates it. His advice to
students recurs all session: <strong>“make yourself bionic with AI”</strong> —
he’d rather hire someone who delivers “abnormal, bionic value” than someone
with a pedigree who refuses the tools.
</p>
<h2 id="ch2-token">The token as the atomic unit</h2>
<p>
“The atomic unit of AI is the token.” And tokens are brutally compute-hungry. The
rough cost to generate a single token scales as the model’s parameter count times the
square of the context length:
</p>
<p style="text-align:center;">$$\text{FLOPs per token} \;\approx\; (\text{parameters}) \times (\text{context length})^2$$</p>
<p>
Sunny’s contrast: a database lookup (e.g., Snowflake) is a modest number of compute
cycles; generating one token is “mind-boggling” by comparison — several
orders of magnitude more than any prior computing paradigm. Then <strong>reasoning
models</strong> arrived, far more voracious in token consumption even before agents — and
token-consumption curves “went parabolic,” with clouds starting to break when
OpenAI and Anthropic each had only ~1 GW of compute.
</p>
<h2 id="ch2-groq">Groq and deterministic compute</h2>
<p>
Groq was founded by <strong>Jonathan Ross</strong> — creator of Google’s TPU, a
high-school dropout who went straight into a PhD math program. Groq’s chip is a
<strong>deterministic data-flow architecture</strong>: paired with a compiler that
pre-determines exactly where every calculation happens.
</p>
<div class="table-wrap">
<table>
<thead><tr><th>Dimension</th><th>GPU (Nvidia)</th><th>Groq</th></tr></thead>
<tbody>
<tr><td>Memory</td><td>Lots of external HBM (slower)</td><td>Lots of on-chip <strong>SRAM</strong> (>10× faster bandwidth, like a giant L1 cache)</td></tr>
<tr><td>Compute</td><td>Very high on-die compute</td><td>Less compute, memory-bandwidth optimized</td></tr>
<tr><td>Execution</td><td>Dynamic scheduling</td><td>Fully deterministic, compiler-placed</td></tr>
<tr><td>Interconnect</td><td>NVLink (72 GPUs today → 576)</td><td>Custom protocol, thousands of chips (ran models on 4–8K chips)</td></tr>
</tbody>
</table>
</div>
<p>
The go-to-market insight: rather than convince people to buy unfamiliar hardware, Groq put
chips <em>in the cloud</em> behind an API and served the best open-source models (including
OpenAI’s Whisper). Developers are fungible if there’s an API — Groq Cloud went
from a couple hundred thousand to <strong>~4M users in a short time</strong> (Nvidia took ~17
years to reach ~7M). Notably, a 2018/2019-era <strong>14nm</strong> Groq chip stayed
competitive against Hopper — five silicon generations newer.
</p>
<div class="quote-box">
<p>“Jensen came on my podcast and said, ‘Everything just changed — inference-time reasoning. We’ve gone from pre-training to inference-time reasoning, and inference is about to <strong>1 billion X</strong>.’”</p>
<div class="quote-attr">— Brad Gerstner, recounting Jensen Huang</div>
</div>
<h2 id="ch2-deal">The Nvidia–Groq deal: disaggregating inference</h2>
<p>
The technical heart of the deal is <strong>disaggregation</strong>. Inference splits into
<em>prefill</em> and <em>decode</em>; many teams already separate machines for each. Groq went
further and disaggregated <em>within decode</em> — some functions are compute-intensive,
others memory-bandwidth-intensive. Via <strong>NVLink Fusion</strong>, Groq’s SRAM
chips speak to Nvidia GPUs, letting each system run the part of the problem it’s best at.
</p>
<div class="tip-box">
<div class="tip-label">💡 THE FACTORY MENTAL MODEL</div>
<p>An inference data center is a factory: <strong>power + chips go in one end, tokens come out the other.</strong> For the <em>same</em> power-and-building footprint, combining the two architectures yields <strong>~2.5× more tokens</strong>. In a world constrained by power and memory, doubling/tripling tokens per footprint is an enormous economic win for an OpenAI or Anthropic.</p>
</div>
<p>
The deal also reads as a lesson in how big outcomes turn on one decision: Sunny texted Brad an
idea, Brad spent “political capital” emailing Jensen, a working prototype followed,
and ~30 days later came the ~$20B acquisition. It worked culturally <em>because</em> Groq
wasn’t “a better GPU” — an SRAM/deterministic/compiler chip is
complementary, not competitive, to Nvidia’s engineering org. (Nvidia is now “seven
chips across five racks,” already disaggregating storage, CPU, compute, networking.)
</p>
<h2 id="ch2-cost">Why inference keeps getting cheaper (and chips get pricier)</h2>
<div class="stat-grid">
<div class="stat-card"><div class="stat-num">~90%</div><div class="stat-label">Drop in inference cost over the last year</div></div>
<div class="stat-card"><div class="stat-num">~99%</div><div class="stat-label">Drop over the last ~2.5 years</div></div>
</div>
<p>Three inputs drive the unit cost of intelligence:</p>
<ol>
<li><strong>Supply chain</strong> — centered on Taiwan (TSMC), packaging and lithography.</li>
<li><strong>Engineering innovation</strong> — circuit layout, memory bandwidth, quantization (e.g., NVFP4). This is where most gains now come from.</li>
<li><strong>Power</strong> — how much energy you can bring to bear.</li>
</ol>
<div class="warn-box">
<div class="warn-label">⚠️ THE CUBE THAT WON’T SIT STILL</div>
<p>Lithography is hitting limits (Moore’s law slowing), so the industry compensates with bigger chips (Cerebras’ pizza-box wafer) and quantization. But as fast as efficiency improves, <strong>models get bigger</strong> (rumored 1–10 trillion parameters) and <strong>demand keeps rising</strong>. Even a 50× efficiency gain over 5 years can be outrun — which is why <strong>H100 prices are actually going up</strong>, not down.</p>
</div>
<p>
The payoff: OpenAI and Anthropic started with <em>highly negative</em> gross margins
(“produce a widget for $1, sell it for 20¢”). The bet was that inference cost
would fall <em>and</em> willingness-to-pay would rise as intelligence got more valuable. AI is
moving from the “answers” era (autocomplete, a better Google) into the
<strong>“action” era</strong> — agents that build apps, resolve tickets, book
travel. When AI <em>does things</em>, tokens consumed jump ~10× but value delivered jumps
~100×, so willingness-to-pay rises faster than cost. Both labs are now at
“very positive gross margins.”
</p>
<div class="quote-box">
<p>“Technology is highly deflationary — but I’ve never seen something this deflationary this quickly. It’s not a single chip, it’s a factory, with all sorts of Moore’s laws playing out combinatorially across it.”</p>
<div class="quote-attr">— Brad Gerstner</div>
</div>
<h2 id="ch2-bubble">The bubble debate</h2>
<p>
The 2026 question, pressed by skeptics like Bill Gurley: are the labs spending at rates they
can never pay off? Brad’s “Oppenheimer moment” was watching Anthropic’s
revenue scale on the same exponential as intelligence:
</p>
<div class="table-wrap">
<table>
<thead><tr><th>Month</th><th>Anthropic monthly revenue (annualized run-rate)</th></tr></thead>
<tbody>
<tr><td>January</td><td>~$3.5B</td></tr>
<tr><td>February</td><td>~$8B</td></tr>
<tr><td>March</td><td>~$10.5B — <em>+$10B annualized added in a single month</em> (≈ Databricks + Palantir combined)</td></tr>
</tbody>
</table>
</div>
<p>
Crucially, that revenue wasn’t bought with a million salespeople — millions of
self-interested customers independently decided “I have to have this.” Brad
contrasts this with his uncomfortable BG2 exchange with Sam Altman: <strong>~$1.4T in spending
commitments against ~$13B of revenue</strong>; Sam’s reply (“if you don’t
like your investment, I’ll buy back your shares”) wasn’t the reassurance Brad
hoped for — but the revenue curve since then resolved much of the doubt.
</p>
<div class="warn-box">
<div class="warn-label">⚠️ WHY THE INSIDERS AREN’T WORRIED ABOUT OVERBUILD</div>
<p>Today’s capabilities were trained on <em>older</em> hardware — Blackwell/Vera/Rubin-trained models haven’t shipped yet. Anthropic and OpenAI plan to add <strong>more compute this year than all labs combined over the prior decade</strong>, then double again. With the world consuming <strong>tens of trillions of tokens per week</strong>, Sunny & Brad read demand as first/second-inning, not an overbuild.</p>
</div>
<h2 id="ch2-agents">Agents, harnesses, and the safety overhang</h2>
<p>
A big driver of token growth is the <strong>harness</strong> around models — Claude
Code, Codex, and tools like “Open Claw” that keep an agent in a continuous loop,
pinging you on your phone when stuck and working all night. Inside Nvidia, a personal
assistant wired to Slack, Teams, email, and files drafts each morning’s task list; agents
increasingly email <em>other</em> agents. Anecdote: Marc Andreessen noted people now spending
$100–$1,000/day on token consumption.
</p>
<div class="quote-box">
<p>“This week Anthropic’s unreleased model, Mythos, found a bug in BSD — code that countless engineers had reviewed. We’ve gone to a place where it’s doing things beyond human capability. And we’re in year three.”</p>
<div class="quote-attr">— Sunny Madra</div>
</div>
<p>
On safety: a model (“Mythos”) was sandboxed, “tried to escape,” and was
hardened via <strong>Project Glass Wing</strong> — an industry consortium (Amazon,
Microsoft, etc.) that found and patched ~26 Safari vulnerabilities before public release.
Brad’s framing: powerful technology is dual-use (“splitting the atom”) —
be optimistic but not Pollyannaish, and beware fear-mongering aimed at regulatory capture.
</p>
<h3>Careers: IQ commoditizes, EQ appreciates</h3>
<p>
A student asked the question on everyone’s mind: <em>if AI keeps driving the marginal
value of raw intelligence toward zero, how do we make sure we’re not wasting our
time?</em> Brad’s answer is part historical pattern, part bet on which human skills
survive. The pattern: every major shift destroys a category of work and pushes humans to a
higher-order one. The craftsperson who could build a wheel start-to-finish was genuinely great
at it — and was completely disintermediated by mass production; it didn’t feel good,
but the world didn’t stop and those people found new things to do. We went from ~80% of
people in farming and manufacturing to ~70% in <em>services</em> — paying for jobs that
barely existed before (coaches, instructors, therapists). The base rate is that humans
<strong>relocate up the value chain</strong>, not into permanent idleness.
</p>
<div class="vs2">
<div class="vs-card vs-down">
<h4>📉 IQ — commoditizes <span class="vs-tag">cheap & abundant</span></h4>
<div class="vs-sub">What the machine now does instantly, at infinite scale</div>
<ul>
<li>Being the fastest solver at the whiteboard</li>
<li>Recalling more facts / knowing more</li>
<li>Raw computation and analysis speed</li>
<li>One-shot “right answer” problem-solving</li>
</ul>
</div>
<div class="vs-card vs-up">
<h4>📈 EQ — appreciates <span class="vs-tag">scarce & durable</span></h4>
<div class="vs-sub">The human-to-human bottleneck that turns intelligence into outcomes</div>
<ul>
<li><strong>Networks</strong> — who you can mobilize</li>
<li><strong>Persuasion</strong> — changing the mind of the person next to you</li>
<li><strong>Team formation</strong> — assembling & holding a group together</li>
<li><strong>Leadership, taste & accountability</strong> — deciding <em>what</em> is worth doing</li>
</ul>
</div>
</div>
<p>
When a capability becomes abundant and near-free, its price collapses — and you
can’t out-compute a machine at the whiteboard. So scarcity, and value, move to the
relational, judgment-laden work machines don’t do. Brad framed it as a future tweet:
<strong>IQ gets commoditized, EQ becomes super valuable.</strong>
</p>
<div class="tip-box">
<div class="tip-label">💡 THE ACTIONABLE HALF: “MAKE YOURSELF BIONIC”</div>
<p>Brad says he no longer cares much where you went to school — he wants someone who shows up and delivers <strong>“abnormal, bionic value,”</strong> meaning a human who pairs their EQ with the machine’s IQ. The losing move is the candidate who says <em>“I don’t use AI, I do everything by hand.”</em> The nuance: you still must learn the fundamentals — you can’t lead or judge work you don’t understand. EQ isn’t a substitute for competence; it’s the durable layer <em>on top</em> of it.</p>
</div>
<p>
The flip side of commoditized intelligence is <strong>access</strong>. Expertise that used to be
a luxury becomes broadly available — intelligence as a public good that raises the floor
for billions, even as it removes the premium on raw cleverness:
</p>
<div class="stat-grid">
<div class="stat-card"><div class="stat-num">~98%</div><div class="stat-label">Who could never afford a specialized private <strong>tutor</strong> — now can</div></div>
<div class="stat-card"><div class="stat-num">~97%</div><div class="stat-label">Who were priced out of <strong>concierge-grade</strong> medicine / advice — now reachable</div></div>
</div>
<p>
Sunny added an optimistic coda: with machines now making discoveries on their own — novel
mathematics, without waiting for “an apple to fall on Newton’s head” — the
pace of progress goes more vertical, which only <em>raises</em> the premium on humans who can
direct and harness it. It rhymes with <strong>Ali Ghodsi’s</strong> point in Chapter 4:
the binding constraint isn’t model intelligence, it’s the human work of context,
judgment, and organizational rewiring — the EQ layer, again.
</p>
<p>
Brad’s closing forecast: he’s publicly predicted <strong>Nvidia will be the first
$10T company</strong> — trading ~13× earnings, ~70% growth, ~$1T of sales booked
over the next eight quarters — arguing the bull case is less about one company than about
the sheer size of the market for intelligence. Apple, by contrast, is “nervous”
about its wait-and-ride strategy, constrained by privacy and the fact that even an 8B-parameter
model can drain an iPhone in 30 minutes.
</p>
</div>
<div class="chapter-footer">
<button class="btn-collapse" onclick="toggleChapter('ch2')">⌃ Collapse</button>
<button class="btn-next" onclick="scrollToChapter('ch3')">Next Chapter →</button>
</div>
</div>
<!-- ════════════════ CHAPTER 3 ════════════════ -->
<div class="chapter" id="ch3">
<div class="chapter-header" onclick="toggleChapter('ch3')">
<div class="chapter-header-left">
<div class="chapter-num">3</div>
<div class="chapter-title-group">
<div class="chapter-title">🏭 From Electrons to Tokens</div>
<div class="chapter-label">Class 3 · Chase Lochmiller (Crusoe)</div>
</div>
</div>
<span class="chevron">⌃</span>
</div>
<div class="chapter-body">
<span class="speaker-tag">Guest: Chase Lochmiller · Founder/CEO, Crusoe</span>
<p>
“Today we’re going to talk about the data centers that you guys are melting.”
Class 3 makes the build-out physical. Chase founded Crusoe, a vertically-integrated AI
infrastructure company, and is an avid mountaineer — the company’s ethos is
“plan A, plan B, and a plan C to the plan B.” The deck’s title:
<strong>from electrons to tokens.</strong>
</p>
<h2 id="ch3-capex">The CapEx supercycle</h2>
<p>
The defining chart is hyperscaler AI CapEx going “up and to the right, fast.”
To contextualize the magnitude:
</p>
<div class="tip-box">
<div class="tip-label">💡 SCALE OF THE BUILD-OUT</div>
<p>This is <strong>bigger than the space program, bigger than the highway system, bigger than the Manhattan Project</strong> — second only to the US defense budget. The hyperscalers are spending on the order of <strong>$650B</strong> building these data centers, and many of those companies are Crusoe customers.</p>
</div>
<p>What does it actually take to produce AI? Chase’s basic equation:</p>
<p style="text-align:center;">$$\text{AI} \;=\; \text{Data} \;+\; \text{Algorithms} \;+\; \text{Compute} \;+\; \text{Energy} \;+\; \text{Data Centers}$$</p>
<ul>
<li><strong>Data</strong> — created a market for labeling (Scale AI, Mercor, even Handshake).</li>
<li><strong>Algorithms</strong> — backprop, transformers, recursive learning; lives in the labs.</li>
<li><strong>Compute + Energy + Data Centers</strong> — Crusoe’s focus, and “where a lot of the money’s being spent.”</li>
</ul>
<h2 id="ch3-diglabor">Why tokens are valuable: digital labor</h2>
<p>
Chase reaches for the <strong>Cobb-Douglas</strong> production model — GDP growth is a
function of changes in labor, capital, and technology:
</p>
<p style="text-align:center;">$$\Delta\,\text{GDP} \;=\; f\big(\Delta L,\; \Delta K,\; \Delta T\big)$$</p>
<p>
Historically, labor (ΔL) could only grow via the birth rate — a 20-year lead
time with a massive incubation period. <strong>For the first time in history we can grow
labor digitally</strong>: when you hand an agent a task (“build me a CRM”), that
is digital labor brought into the world by buying GPUs and data centers. That’s why the
investment is so broad-based — the prize is accelerating GDP itself.
</p>
<h2 id="ch3-energy">The energy-first thesis</h2>
<p>
Crusoe is vertically integrated specifically because <strong>the bottleneck keeps
moving</strong>: from chips, to power, to memory, to labor, to electrical equipment (switchgear,
chillers, turbines). Today the binding constraint is <strong>energized, powered shells</strong>
— places where you can plug in chips and turn them on. Chips themselves have
“softened” as a bottleneck.
</p>
<p>
Crusoe’s founding insight (the “Robinson Crusoe” name): markets are roughly
efficient, so don’t be the next me-too developer in Northern Virginia. Instead, find
stranded low-cost energy and <strong>move the data to the power, not the power to the
data.</strong>